diff --git a/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp b/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp
index 8424d0f..e7a38a9 100644
--- a/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp
+++ b/libomptarget/plugins-nextgen/amdgpu/src/rtl.cpp
@@ -1976,12 +1976,18 @@
 
   virtual Error callGlobalConstructors(GenericPluginTy &Plugin,
                                        DeviceImageTy &Image) override {
-    return callGlobalCtorDtorCommon(Plugin, Image, "amdgcn.device.init");
+    GenericGlobalHandlerTy &Handler = Plugin.getGlobalHandler();
+    if (Handler.isSymbolInImage(*this, Image, "amdgcn.device.fini"))
+      Image.setPendingGlobalDtors();
+
+    return callGlobalCtorDtorCommon(Plugin, Image, /*IsCtor=*/true);
   }
 
   virtual Error callGlobalDestructors(GenericPluginTy &Plugin,
                                       DeviceImageTy &Image) override {
-    return callGlobalCtorDtorCommon(Plugin, Image, "amdgcn.device.fini");
+    if (Image.hasPendingGlobalDtors())
+      return callGlobalCtorDtorCommon(Plugin, Image, /*IsCtor=*/false);
+    return Plugin::success();
   }
 
   const uint64_t getStreamBusyWaitMicroseconds() const {
@@ -2701,15 +2707,17 @@
   /// Common method to invoke a single threaded constructor or destructor
   /// kernel by name.
   Error callGlobalCtorDtorCommon(GenericPluginTy &Plugin, DeviceImageTy &Image,
-                                 const char *Name) {
+                                 bool IsCtor) {
+    const char *KernelName =
+        IsCtor ? "amdgcn.device.init" : "amdgcn.device.fini";
     // Perform a quick check for the named kernel in the image. The kernel
     // should be created by the 'amdgpu-lower-ctor-dtor' pass.
     GenericGlobalHandlerTy &Handler = Plugin.getGlobalHandler();
-    if (!Handler.isSymbolInImage(*this, Image, Name))
+    if (IsCtor && !Handler.isSymbolInImage(*this, Image, KernelName))
       return Plugin::success();
 
     // Allocate and construct the AMDGPU kernel.
-    AMDGPUKernelTy AMDGPUKernel(Name);
+    AMDGPUKernelTy AMDGPUKernel(KernelName);
     if (auto Err = AMDGPUKernel.init(*this, Image))
       return Err;
 
diff --git a/libomptarget/plugins-nextgen/common/include/PluginInterface.h b/libomptarget/plugins-nextgen/common/include/PluginInterface.h
index b85dc14..f7f7236 100644
--- a/libomptarget/plugins-nextgen/common/include/PluginInterface.h
+++ b/libomptarget/plugins-nextgen/common/include/PluginInterface.h
@@ -218,12 +218,19 @@
   const __tgt_device_image *TgtImage;
   const __tgt_device_image *TgtImageBitcode;
 
+  /// If this image has any global destructors that much be called.
+  /// FIXME: This is only required because we currently have no invariants
+  ///        towards the lifetime of the underlying image. We should either copy
+  ///        the image into memory locally or erase the pointers after init.
+  bool PendingGlobalDtors;
+
   /// Table of offload entries.
   OffloadEntryTableTy OffloadEntryTable;
 
 public:
   DeviceImageTy(int32_t Id, const __tgt_device_image *Image)
-      : ImageId(Id), TgtImage(Image), TgtImageBitcode(nullptr) {
+      : ImageId(Id), TgtImage(Image), TgtImageBitcode(nullptr),
+        PendingGlobalDtors(false) {
     assert(TgtImage && "Invalid target image");
   }
 
@@ -255,6 +262,10 @@
                            "Image");
   }
 
+  /// Accessors to the boolean value
+  bool setPendingGlobalDtors() { return PendingGlobalDtors = true; }
+  bool hasPendingGlobalDtors() const { return PendingGlobalDtors; }
+
   /// Get a reference to the offload entry table for the image.
   OffloadEntryTableTy &getOffloadEntryTable() { return OffloadEntryTable; }
 };
diff --git a/libomptarget/plugins-nextgen/cuda/src/rtl.cpp b/libomptarget/plugins-nextgen/cuda/src/rtl.cpp
index 7ccd0b4..ce6b398 100644
--- a/libomptarget/plugins-nextgen/cuda/src/rtl.cpp
+++ b/libomptarget/plugins-nextgen/cuda/src/rtl.cpp
@@ -392,12 +392,20 @@
 
   virtual Error callGlobalConstructors(GenericPluginTy &Plugin,
                                        DeviceImageTy &Image) override {
+    // Check for the presense of global destructors at initialization time. This
+    // is required when the image may be deallocated before destructors are run.
+    GenericGlobalHandlerTy &Handler = Plugin.getGlobalHandler();
+    if (Handler.isSymbolInImage(*this, Image, "nvptx$device$fini"))
+      Image.setPendingGlobalDtors();
+
     return callGlobalCtorDtorCommon(Plugin, Image, /*IsCtor=*/true);
   }
 
   virtual Error callGlobalDestructors(GenericPluginTy &Plugin,
                                       DeviceImageTy &Image) override {
-    return callGlobalCtorDtorCommon(Plugin, Image, /*IsCtor=*/false);
+    if (Image.hasPendingGlobalDtors())
+      return callGlobalCtorDtorCommon(Plugin, Image, /*IsCtor=*/false);
+    return Plugin::success();
   }
 
   Expected<std::unique_ptr<MemoryBuffer>>
@@ -1145,7 +1153,7 @@
     // Perform a quick check for the named kernel in the image. The kernel
     // should be created by the 'nvptx-lower-ctor-dtor' pass.
     GenericGlobalHandlerTy &Handler = Plugin.getGlobalHandler();
-    if (!Handler.isSymbolInImage(*this, Image, KernelName))
+    if (IsCtor && !Handler.isSymbolInImage(*this, Image, KernelName))
       return Plugin::success();
 
     // The Nvidia backend cannot handle creating the ctor / dtor array