diff --git a/scripts/vllm_infer.py b/scripts/vllm_infer.py
index 4cdc290c..2b78fa97 100644
--- a/scripts/vllm_infer.py
+++ b/scripts/vllm_infer.py
@@ -45,6 +45,7 @@ def vllm_infer(
     top_k: int = 50,
     max_new_tokens: int = 1024,
     repetition_penalty: float = 1.0,
+    seed: int = None,
     pipeline_parallel_size: int = 1,
     image_max_pixels: int = 768 * 768,
     image_min_pixels: int = 32 * 32,
@@ -108,6 +109,7 @@ def vllm_infer(
         stop_token_ids=template_obj.get_stop_token_ids(tokenizer),
         max_tokens=generating_args.max_new_tokens,
         skip_special_tokens=False,
+        seed=seed,
     )
     if model_args.adapter_name_or_path is not None:
         lora_request = LoRARequest("default", 1, model_args.adapter_name_or_path[0])