diff --git a/python/sglang/srt/models/deepseek_v2.py b/python/sglang/srt/models/deepseek_v2.py
index 5399d6904..aa47daa30 100644
--- a/python/sglang/srt/models/deepseek_v2.py
+++ b/python/sglang/srt/models/deepseek_v2.py
@@ -462,7 +462,7 @@ class DeepseekV2MoE(nn.Module):
         if hasattr(self, "shared_experts") and use_intel_amx_backend(
             self.shared_experts.gate_up_proj
         ):
-            return self.forward_cpu(hidden_states)
+            return self.forward_cpu(hidden_states, can_fuse_mlp_allreduce)
 
         shared_output = self._forward_shared_experts(hidden_states)
         # router_logits: (num_tokens, n_experts)
@@ -479,7 +479,9 @@ class DeepseekV2MoE(nn.Module):
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states
 
-    def forward_cpu(self, hidden_states: torch.Tensor) -> torch.Tensor:
+    def forward_cpu(
+        self, hidden_states: torch.Tensor, can_fuse_mlp_allreduce: bool = False
+    ) -> torch.Tensor:
         # router_logits: (num_tokens, n_experts)
         router_logits = self.gate(hidden_states)
         fused_experts_out = self.experts(
@@ -528,7 +530,7 @@ class DeepseekV2MoE(nn.Module):
             None,  # a2_scale
             True,  # is_vnni
         )
-        if self.tp_size > 1 and not self.can_fuse_mlp_allreduce:
+        if self.tp_size > 1 and not can_fuse_mlp_allreduce:
             final_hidden_states = tensor_model_parallel_all_reduce(final_hidden_states)
         return final_hidden_states