diff --git a/src/arch/cuda/hybridAPI/hapi_impl.cpp b/src/arch/cuda/hybridAPI/hapi_impl.cpp
index beeea38d24..309bd6e23b 100644
--- a/src/arch/cuda/hybridAPI/hapi_impl.cpp
+++ b/src/arch/cuda/hybridAPI/hapi_impl.cpp
@@ -1367,6 +1367,7 @@ void hapiPollEvents(void* param) {
   std::queue<hapiEvent>& queue = CpvAccess(hapi_event_queue);
   while (!queue.empty()) {
     hapiEvent hev = queue.front();
+    queue.pop(); // TODO: investigate possible race condition with charm4py futures - temporarily resolved by popping here
     if (cudaEventQuery(hev.event) == cudaSuccess) {
       // invoke Charm++ callback if one was given
       hev.cb.send(hev.cb_msg);
@@ -1376,7 +1377,6 @@ void hapiPollEvents(void* param) {
         hapiWorkRequestCleanup(hev.wr);
       }
       cudaEventDestroy(hev.event);
-      queue.pop();
       CpvAccess(n_hapi_events)--;
 
       // inform QD that an event was processed
diff --git a/src/ck-core/charm.h b/src/ck-core/charm.h
index cee65d539a..b645027626 100644
--- a/src/ck-core/charm.h
+++ b/src/ck-core/charm.h
@@ -510,6 +510,11 @@ extern int CkDisableTracing(int epIdx);
 extern void CkEnableTracing(int epIdx);
 extern void CkCallWhenIdle(int epIdx, void* obj);
 
+
+#if CMK_CHARM4PY && CMK_CUDA
+extern void CkHapiAddCallback(long stream, void (*cb)(void*, void*), int fid);
+#endif
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/src/ck-core/ck.C b/src/ck-core/ck.C
index 33bfc82bcf..8e6aa5e470 100644
--- a/src/ck-core/ck.C
+++ b/src/ck-core/ck.C
@@ -2637,6 +2637,18 @@ void CkArrayExtSend_multi(int aid, int *idx, int ndims, int epIdx, int num_bufs,
   }
 }
 
+
+#if CMK_CUDA
+#include "hapi.h"
+void CkHapiAddCallback(long stream, void (*cb)(void*, void*), int fid) 
+{
+  cudaStream_t stream_ptr = (cudaStream_t)stream;
+  CkCallback callback(cb, (void *) fid);
+  
+  hapiAddCallback(stream_ptr, callback, NULL);
+}
+#endif // CMK_CUDA
+
 #endif
 
 //------------------- Message Watcher (record/replay) ----------------