diff --git a/fastdeploy/model_executor/layers/backends/xpu/moe/fused_moe.py b/fastdeploy/model_executor/layers/backends/xpu/moe/fused_moe.py
index a7ac1845a21..4636d82ddff 100644
--- a/fastdeploy/model_executor/layers/backends/xpu/moe/fused_moe.py
+++ b/fastdeploy/model_executor/layers/backends/xpu/moe/fused_moe.py
@@ -262,7 +262,7 @@ def apply(
         #     layer.top_k,
         #     False,  # moe group, used in deepseek
         # )
-        # if layer.tp_size > 1:
+        # if layer.reduce_results and layer.tp_size > 1:
         #     from fastdeploy.distributed.communication import (
         #         tensor_model_parallel_all_reduce,
         #     )
@@ -271,48 +271,53 @@ def apply(
 
         # return fused_moe_out
 
-        gate_out = paddle.matmul(x.cast("float32"), gate.weight.transpose([1, 0]), transpose_y=True)
-        topk_idx, topk_weights = moe_topk_select(gate_out, layer.gate_correction_bias, layer.top_k, True)
-        token_nums_per_expert_list = list(range(64))  # 填充做占位符
-        permute_input, permute_indices_per_token, token_num_lod, dst_weights, ffn1_act_scale_per_token = (
-            ep_moe_expert_dispatch(
-                x,
-                topk_idx,
-                topk_weights,
-                (layer.up_gate_proj_in_scale if hasattr(layer, "up_gate_proj_in_scale") else None),
-                token_nums_per_expert_list,
-                x.shape[0] * layer.top_k,
+        token_num = x.shape[0]
+        if token_num > 0:
+            gate_out = paddle.matmul(x.cast("float32"), gate.weight.transpose([1, 0]), transpose_y=True)
+            topk_idx, topk_weights = moe_topk_select(gate_out, layer.gate_correction_bias, layer.top_k, True)
+            token_nums_per_expert_list = list(range(64))  # 填充做占位符
+            permute_input, permute_indices_per_token, token_num_lod, dst_weights, ffn1_act_scale_per_token = (
+                ep_moe_expert_dispatch(
+                    x,
+                    topk_idx,
+                    topk_weights,
+                    (layer.up_gate_proj_in_scale if hasattr(layer, "up_gate_proj_in_scale") else None),
+                    token_nums_per_expert_list,
+                    x.shape[0] * layer.top_k,
+                    self.moe_quant_type,
+                )
+            )
+
+            ffn_out = moe_expert_ffn(
+                permute_input,
+                token_num_lod,
+                layer.up_gate_proj_weight,
+                layer.down_proj_weight,
+                None,  # moe_ffn1_bias
+                None,  # moe_ffn2_bias
+                None,  # ffn1 in scale
+                None,  # ffn2 in scale
+                (layer.up_gate_proj_weight_scale if hasattr(layer, "up_gate_proj_weight_scale") else None),
+                (layer.down_proj_weight_scale if hasattr(layer, "down_proj_weight_scale") else None),
+                None,  # moe_ffn2_shift
+                None,  # moe_ffn2_smooth
                 self.moe_quant_type,
+                -1,
+                x.shape[0] * layer.top_k,  # token_all_num
             )
-        )
+            topk_weights_bf16 = topk_weights.astype("bfloat16")
+            tmp_ffn_out = ep_moe_expert_combine(
+                ffn_out,
+                permute_indices_per_token,
+                topk_weights_bf16,
+                permute_indices_per_token.shape[0],
+                ffn_out.shape[0],
+                ffn_out.shape[1],
+                permute_indices_per_token.shape[1],
+            )
+        else:
+            tmp_ffn_out = paddle.empty(x.shape, x.dtype)
 
-        ffn_out = moe_expert_ffn(
-            permute_input,
-            token_num_lod,
-            layer.up_gate_proj_weight,
-            layer.down_proj_weight,
-            None,  # moe_ffn1_bias
-            None,  # moe_ffn2_bias
-            None,  # ffn1 in scale
-            None,  # ffn2 in scale
-            (layer.up_gate_proj_weight_scale if hasattr(layer, "up_gate_proj_weight_scale") else None),
-            (layer.down_proj_weight_scale if hasattr(layer, "down_proj_weight_scale") else None),
-            None,  # moe_ffn2_shift
-            None,  # moe_ffn2_smooth
-            self.moe_quant_type,
-            -1,
-            x.shape[0] * layer.top_k,  # token_all_num
-        )
-        topk_weights_bf16 = topk_weights.astype("bfloat16")
-        tmp_ffn_out = ep_moe_expert_combine(
-            ffn_out,
-            permute_indices_per_token,
-            topk_weights_bf16,
-            permute_indices_per_token.shape[0],
-            ffn_out.shape[0],
-            ffn_out.shape[1],
-            permute_indices_per_token.shape[1],
-        )
         if layer.reduce_results and layer.tp_size > 1:
             from fastdeploy.distributed.communication import (
                 tensor_model_parallel_all_reduce,
@@ -858,7 +863,7 @@ def apply(
             ffn_out.shape[1],
             permute_indices_per_token.shape[1],
         )
-        if layer.tp_size > 1:
+        if layer.reduce_results and layer.tp_size > 1:
             from fastdeploy.distributed.communication import (
                 tensor_model_parallel_all_reduce,
             )