[fix]解决deepseek模型cudagraph模式下精度异常问题

vllm/model_executor/models/deepseek_v2.py

@@ -181,8 +181,13 @@ class DeepseekV2MoE(nn.Module):
                                      quant_config=None,
                                      prefix=f"{prefix}.gate")
         if config.topk_method == "noaux_tc":
-            self.gate.e_score_correction_bias = nn.Parameter(
-                torch.empty(config.n_routed_experts, dtype=torch.float32))
+            if envs.VLLM_ENABLE_MOE_FUSED_GATE:
+                # avoid moe_fused_gate precision error
+                self.gate.e_score_correction_bias = nn.Parameter(
+                torch.empty(config.n_routed_experts))
+            else:
+                self.gate.e_score_correction_bias = nn.Parameter(
+                    torch.empty(config.n_routed_experts, dtype=torch.float32))
         else:
             self.gate.e_score_correction_bias = None