Make pp_stream wait on attn_backward_dx

Author: lshpku

paddlenlp/transformers/deepseek_v2/modeling_pp.py

@@ -1077,7 +1077,7 @@ def forward_backward(self, inputs, output_grad, combine_bw_event_to_wait=None, p
         # get dispatch backward event
         dispatch_backward_event = deep_ep.get_event_from_comm_stream(self.backward_node.moe_group.id)
 
-        paddle.base.core.nvprof_nvtx_push("dispatch_backward_dw")
+        paddle.base.core.nvprof_nvtx_push("mlp_backward_dw")
         WeightGradStore.pop()
         assert WeightGradStore.funcs_queue.empty()
         paddle.base.core.nvprof_nvtx_pop()
@@ -1108,11 +1108,7 @@ def forward_backward(self, inputs, output_grad, combine_bw_event_to_wait=None, p
 
         if pp_stream is not None:
             send_recv_stream = paddle.device.Stream(stream_base=pp_stream)
-
-            # combine_forward_event.custom_stream_wait( pp_stream)
-            # final_out_event.custom_stream_wait(pp_stream)
-
-            paddle.base.core.nvprof_nvtx_push("pp stream add")
+            paddle.base.core.nvprof_nvtx_push("pp_stream_add")
 
             with paddle.device.stream_guard(send_recv_stream):
                 combine_forward_event.current_stream_wait()
@@ -1127,20 +1123,27 @@ def forward_backward(self, inputs, output_grad, combine_bw_event_to_wait=None, p
 
         dispatch_backward_event.calc_stream_wait(self.backward_node.moe_group.id)
 
-        paddle.base.core.nvprof_nvtx_push("attn_backward")
+        paddle.base.core.nvprof_nvtx_push("attn_backward_dx")
         assert WeightGradStore.funcs_queue.empty()
         WeightGradStore.enabled = True
         output_grad = self.backward_node.attn_backward(output_grad)
         event_to_wait = deep_ep.get_event_from_calc_stream(self.backward_node.moe_group.id)
+        paddle.base.core.nvprof_nvtx_pop()
 
         if EventStore is not None:
             EventStore.set(event_to_wait)
 
+        if pp_stream is not None:
+            # TODO(liangshuhao): this wait may be unnecessary, but we would suffer slow
+            # convergence rate without this, so we temporarily put a wait here.
+            with paddle.device.stream_guard(send_recv_stream):
+                event_to_wait.current_stream_wait()
+
+        paddle.base.core.nvprof_nvtx_push("attn_backward_dw")
         WeightGradStore.enabled = False
         WeightGradStore.flush()
         WeightGradStore.pop()
         assert WeightGradStore.funcs_queue.empty()
-
         paddle.base.core.nvprof_nvtx_pop()
 
         # residual add