diff --git a/vllm_ascend/worker/model_runner_v1.py b/vllm_ascend/worker/model_runner_v1.py
index 923ff0c..5a95147 100644
--- a/vllm_ascend/worker/model_runner_v1.py
+++ b/vllm_ascend/worker/model_runner_v1.py
@@ -953,6 +953,7 @@ class NPUModelRunner(LoRAModelRunnerMixin):
                 self.mrope_positions_cpu[:, :total_num_scheduled_tokens],
                 non_blocking=True)
 
+        self.positions[total_num_scheduled_tokens:num_input_tokens].zero_()
         self.positions[:total_num_scheduled_tokens].copy_(
             self.positions_cpu[:total_num_scheduled_tokens], non_blocking=True)
         positions = self.positions[:num_input_tokens]