diff --git a/paddle/fluid/platform/device_event_base.h b/paddle/fluid/platform/device_event_base.h
index d0458dcb9e4e46ba4836e862264db0d36b048b31..f56688de09a326e6846d8d423fbaf66e9c0f6e9f 100644
--- a/paddle/fluid/platform/device_event_base.h
+++ b/paddle/fluid/platform/device_event_base.h
@@ -85,6 +85,9 @@ class DeviceEvent {
         event_recorder_[type_id_],
         platform::errors::Unavailable(
             "event_recorder_[%d] shall not be nullptr.", type_id_));
+    if (!recorded_) {
+      recorded_ = true;
+    }
     event_recorder_[type_id_](this, dev_ctx);
   }
 
@@ -93,6 +96,10 @@ class DeviceEvent {
         event_querier_[type_id_],
         platform::errors::Unavailable(
             "event_querier_[%d] shall not be nullptr.", type_id_));
+    if (!recorded_) {
+      VLOG(4) << "Event " << this << " is not recorded yet, and skip query!";
+      return true;
+    }
     return event_querier_[type_id_](this);
   }
 
@@ -127,6 +134,10 @@ class DeviceEvent {
                                 "event_waiter_[%d][%d] shall not be nullptr.",
                                 waiter_idx,
                                 type_id_));
+    if (!recorded_) {
+      VLOG(4) << "Event " << this << " is not recorded yet, and skip wait!";
+      return;
+    }
     event_waiter_[waiter_idx][type_id_](this, context);
   }
 
@@ -140,6 +151,14 @@ class DeviceEvent {
   int type_id_;
   unsigned int flag_;
 
+  // NOTE(chenruibiao): In cross-step stream synchronization, an event may be
+  // recorded in the first step and waited in the second step. So, in the first
+  // step, the WaitEvent may be called without RecordEvent.
+  // On cuda device, it is ok to wait event that is not recorded yet;
+  // while on npu device, it results in error.
+  // So, we add flag recorded_ to handle this case uniformly.
+  bool recorded_{false};
+
   static EventCreateFunction event_creator_[MaxDeviceTypes];
   static EventRecordFunction event_recorder_[MaxDeviceTypes];
   static EventQueryFunction event_querier_[MaxDeviceTypes];