fix bug (#10856)

Author: zhangbo9674

paddlenlp/transformers/deepseek_v2/modeling_pp.py

@@ -629,7 +629,7 @@ def combine_backward(self, output_grad, async_finish=False, allocate_on_comm_str
             output_combine_grad,
             async_finish=async_finish,
             previous_event=quant_event,
-            allocate_on_comm_stream=allocate_on_comm_stream,
+            allocate_on_comm_stream=allocate_on_comm_stream and quant_event is not None,
         )
 
         ret = (hidden_states_grad, residual_grad, l_aux_grad, hidden_states_out_grad)

paddlenlp/transformers/moe_utils.py

@@ -160,6 +160,7 @@ def forward(
                 num_experts=num_experts,
                 tokens_per_expert=tokens_per_expert,
                 padding_multiplex=128,
+                fill_output=True,
             )
         else:
             with paddle.amp.auto_cast(False):
@@ -236,6 +237,7 @@ def backward(
                 num_experts,
                 tokens_per_expert,
                 padding_multiplex=128,
+                fill_output=True,
             )
             return (unzipped_grad, unzipped_scale_grad)
         else: