diff --git a/src/gpu/dpcpp_ccl.cpp b/src/gpu/dpcpp_ccl.cpp
index 3bd8087..c5b5ce3 100644
--- a/src/gpu/dpcpp_ccl.cpp
+++ b/src/gpu/dpcpp_ccl.cpp
@@ -689,7 +689,8 @@ c10::intrusive_ptr<ProcessGroupCCL::AsyncWorkCCL> XPUCCLStubs::allreduce_(std::v
                                             stream,
                                             attr), stream.get_native());
       });
-    // printf("Use One CCL allreduce.\n");
+    stream.get_native().wait();
+   // printf("Use One CCL allreduce.\n");
     return ret_evt;
   },
   c10d::OpType::ALLREDUCE);