update deepseek_v2.py

vllm/model_executor/models/deepseek_v2.py

@@ -257,9 +257,6 @@ class DeepseekV2MoE(nn.Module):
                 reduce_results=False,
                 prefix=f"{prefix}.shared_experts",
             )
-        from vllm.two_batch_overlap.two_batch_overlap import tbo_all_reduce
-        self.tbo_all_reduce = tbo_all_reduce
-
             self.experts = SharedFusedMoE(
                 shared_experts=self.shared_experts,
                 num_experts=config.n_routed_experts,
@@ -281,6 +278,9 @@ class DeepseekV2MoE(nn.Module):
                 num_redundant_experts=self.n_redundant_experts,
                 is_sequence_parallel=self.is_sequence_parallel,
             )
+        from vllm.two_batch_overlap.two_batch_overlap import tbo_all_reduce
+        self.tbo_all_reduce = tbo_all_reduce
+
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         num_tokens, hidden_dim = hidden_states.shape