diff --git a/custom_ops/gpu_ops/speculate_decoding/speculate_limit_thinking_content_length_v2.cu b/custom_ops/gpu_ops/speculate_decoding/speculate_limit_thinking_content_length_v2.cu
index 709911d2ba0..8d963eb0c36 100644
--- a/custom_ops/gpu_ops/speculate_decoding/speculate_limit_thinking_content_length_v2.cu
+++ b/custom_ops/gpu_ops/speculate_decoding/speculate_limit_thinking_content_length_v2.cu
@@ -42,7 +42,7 @@ __global__ void speculate_limit_thinking_content_length_kernel_v2(
   if (max_think_len < 0) return;
   int current_limit_think_status = limit_think_status[bid];
   // 如果在回复阶段, 且已经触发停止标志, 则直接返回, 无需多余执行.
-  if (current_limit_think_status == 3 && stop_flags[bid]) {
+  if (current_limit_think_status == 3 || stop_flags[bid]) {
     return;
   }