修复部分代码

vllm/model_executor/models/deepseek_v2.py

@@ -227,15 +227,21 @@ class DeepseekV2MoE(nn.Module):
         router_logits, _ = self.gate(hidden_states)
 
         if not self.use_mori_ep:
-            if hidden_states.dtype != torch.float16:
+            if envs.VLLM_USE_LIGHTOP and not self.dpsk_fp16_quick:
                 final_hidden_states = self.experts(
                     hidden_states=hidden_states,
-                    router_logits=router_logits) * self.routed_scaling_factor
+                    router_logits=router_logits,
+                    shared_output=shared_output)
             else:
-                # Fix FP16 overflow
-                # See DeepseekV2DecoderLayer for more details.
-                final_hidden_states = self.experts(hidden_states=hidden_states,
-                                                router_logits=router_logits)
+                if hidden_states.dtype != torch.float16 or self.dpsk_fp16_quick:
+                    final_hidden_states = self.experts(
+                        hidden_states=hidden_states,
+                        router_logits=router_logits) * self.routed_scaling_factor
+                else:
+                    # Fix FP16 overflow
+                    # 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)