diff --git a/fastdeploy/worker/gpu_model_runner.py b/fastdeploy/worker/gpu_model_runner.py
index 67a09f5c5d4..ff9156ca9f2 100644
--- a/fastdeploy/worker/gpu_model_runner.py
+++ b/fastdeploy/worker/gpu_model_runner.py
@@ -2387,7 +2387,7 @@ def profile_run(self) -> None:
         # 2. Dummy run
         self._dummy_run(
             num_tokens=self.scheduler_config.max_num_batched_tokens,
-            batch_size=self.scheduler_config.max_num_seqs,
+            batch_size=min(self.scheduler_config.max_num_seqs, 3),
         )
 
         # 3. gc