update deepseek_v2.py

vllm/model_executor/models/deepseek_v2.py

@@ -252,15 +252,6 @@ class DeepseekV2MoE(nn.Module):
             final_hidden_states = self.experts(hidden_states=hidden_states,
                                                 router_logits=router_logits)
 
-        if shared_output is not None:
-            if hidden_states.dtype != torch.float16:
-                final_hidden_states = final_hidden_states + shared_output
-            else:
-                # Fix FP16 overflow
-                # See DeepseekV2DecoderLayer for more details.
-                final_hidden_states = final_hidden_states + shared_output \
-                    * (1. / self.routed_scaling_factor)
-
         if self.tp_size > 1:
             if envs.VLLM_ENABLE_TBO:
                 final_hidden_states = self.tbo_all_reduce(final_hidden_states)