[perf] slightly imporve DeepSeek-R1-FP4 TP8 (#7481)

python/sglang/srt/models/deepseek_v2.py

@@ -362,12 +362,14 @@ class DeepseekV2MoE(nn.Module):
             return self.forward_deepep(hidden_states, forward_batch)
 
     def forward_normal_dual_stream(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # router_logits: (num_tokens, n_experts)
+        router_logits = self.gate(hidden_states)
+
         current_stream = torch.cuda.current_stream()
         self.alt_stream.wait_stream(current_stream)
         shared_output = self._forward_shared_experts(hidden_states)
+
         with torch.cuda.stream(self.alt_stream):
-            # router_logits: (num_tokens, n_experts)
-            router_logits = self.gate(hidden_states)
             final_hidden_states = self.experts(
                 hidden_states=hidden_states, router_logits=router_logits
             )