fix mlp bw split bug (#10932)

Co-authored-by: phlrain <[email protected]>

Author: phlrain

paddlenlp/transformers/deepseek_v2/modeling_pp.py

@@ -870,7 +870,8 @@ def forward_backward(self, inputs, output_grad, combine_bw_event_to_wait=None, p
         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")
-        self.backward_node.mlp_backward_dw()
+        WeightGradStore.pop()
+        assert WeightGradStore.funcs_queue.empty()
         paddle.base.core.nvprof_nvtx_pop()
 
         dispatch_forward_event.calc_stream_wait(self.forward_node.moe_group.id)