diff --git a/lmdeploy/model.py b/lmdeploy/model.py
index 626f3f5b1b..c1700b37d6 100644
--- a/lmdeploy/model.py
+++ b/lmdeploy/model.py
@@ -1583,10 +1583,12 @@ def __init__(self,
                  system='<|system|>\n',
                  user='<|user|>\n',
                  assistant='<|assistant|>\n',
+                 stop_words=['<|user|>', '<|endoftext|>', '<|observation|>'],
                  **kwargs):
         super().__init__(system=system,
                          user=user,
                          assistant=assistant,
+                         stop_words=stop_words,
                          **kwargs)
         self.start = '[gMASK]<sop>'
 
diff --git a/tests/test_lmdeploy/test_model.py b/tests/test_lmdeploy/test_model.py
index f38b952e33..3a00266d19 100644
--- a/tests/test_lmdeploy/test_model.py
+++ b/tests/test_lmdeploy/test_model.py
@@ -329,7 +329,13 @@ def test_chatglm3():
 
 
 def test_glm4():
-    model = MODELS.get('glm4')()
+    model_path_and_name = 'THUDM/glm-4-9b-chat'
+    deduced_name = best_match_model(model_path_and_name)
+    assert deduced_name == 'glm4'
+
+    model = MODELS.get(deduced_name)()
+    # check stop words
+    assert model.stop_words == ['<|user|>', '<|endoftext|>', '<|observation|>']
     messages = [{
         'role': 'system',
         'content': 'you are a helpful assistant'
@@ -344,7 +350,7 @@ def test_glm4():
         'content': 'AGI is?'
     }]
     from transformers import AutoTokenizer
-    tokenizer = AutoTokenizer.from_pretrained('THUDM/glm-4-9b-chat',
+    tokenizer = AutoTokenizer.from_pretrained(model_path_and_name,
                                               trust_remote_code=True)
     ref = tokenizer.apply_chat_template(messages, tokenize=False)
     res = model.messages2prompt(messages)