diff --git a/fastdeploy/model_executor/layers/moe/fused_moe_backend_base.py b/fastdeploy/model_executor/layers/moe/fused_moe_backend_base.py
index efe0ec1be..8c35fa83a 100644
--- a/fastdeploy/model_executor/layers/moe/fused_moe_backend_base.py
+++ b/fastdeploy/model_executor/layers/moe/fused_moe_backend_base.py
@@ -175,12 +175,13 @@ class MoEMethodBase(QuantMethodBase):
         Paddle Cutlass compute Fused MoE.
         """
         if layer.ep_size > 1:
-            if layer.fd_config.model_config.moe_phase.phase == "prefill" and layer.layer_idx == 0:
-                if layer.fd_config.scheduler_config.splitwise_role == "mixed":
+            is_moe_start_layer = layer.layer_idx == layer.fd_config.model_config.moe_layer_start_index
+            if layer.fd_config.model_config.moe_phase.phase == "prefill":
+                if layer.fd_config.scheduler_config.splitwise_role == "mixed" and is_moe_start_layer:
                     self.ep_prefill_runner.clean_low_latency_buffer()
                 return self.apply_ep_prefill(layer, x, gate)
             else:
-                if layer.fd_config.scheduler_config.splitwise_role == "mixed" and layer.layer_idx == 0:
+                if layer.fd_config.scheduler_config.splitwise_role == "mixed" and is_moe_start_layer:
                     self.ep_decoder_runner.clean_low_latency_buffer()
                 return self.apply_ep_decode(layer, x, gate)
         else: