Fix post_process_backward calc stream wait (#11063)

Author: lshpku

paddlenlp/transformers/deepseek_v2/modeling_pp.py

@@ -766,6 +766,9 @@ def post_process_forward(self, inputs, with_residual=True):
         return inputs
 
     def post_process_backward(self, output_grad, event_to_wait=None):
+        if event_to_wait is not None:
+            event_to_wait.calc_stream_wait(self.moe_group.id)
+
         grad = self.post_process_node.backward(output_grad)
 
         if self.using_post_norm_recompute: