fix deepseek_v2.py

vllm/model_executor/models/deepseek_v2.py

@@ -242,11 +242,6 @@ class DeepseekV2MoE(nn.Module):
                     # See DeepseekV2DecoderLayer for more details.
                     final_hidden_states = self.experts(hidden_states=hidden_states,
                                                     router_logits=router_logits)
-        else:        
-            final_hidden_states = self.experts(hidden_states=hidden_states,
-                                                router_logits=router_logits)
-
-        if not self.use_mori_ep:
                 if shared_output is not None:
                     if hidden_states.dtype != torch.float16 or self.dpsk_fp16_quick:
                         final_hidden_states = final_hidden_states + shared_output
@@ -255,7 +250,11 @@ class DeepseekV2MoE(nn.Module):
                         # See DeepseekV2DecoderLayer for more details.
                         final_hidden_states = final_hidden_states + shared_output \
                             * (1. / self.routed_scaling_factor)
+        else:        
+            final_hidden_states = self.experts(hidden_states=hidden_states,
+                                                router_logits=router_logits)
 
+        if not self.use_mori_ep:
             if self.tp_size > 1:
                 if envs.VLLM_ENABLE_TBO:
                     final_hidden_states = self.tbo_all_reduce(final_hidden_states)