fix some errors in inference.py

inference.py

@@ -279,10 +279,10 @@ class Chat:
             top_k=20,
             top_p=0.9,
             temperature=temperature,
-            max_new_tokens=max_new_tokens,
-            stopping_criteria=stopping_criteria
-        )
+            max_new_tokens=max_new_tokens
 
+        )
+        self.stopping_criteria = stopping_criteria
     def ask(self, text, conv, modal_type="image"):
         assert modal_type in ["text", "image", "video"]
         conversations = []
@@ -331,6 +331,7 @@ class Chat:
                 input_ids=input_ids,
                 attention_mask=attention_mask,
                 generation_config=self.generation_config,
+                stopping_criteria=self.stopping_criteria,
                 return_dict_in_generate=True,
                 output_scores=True
             )
@@ -345,6 +346,7 @@ class Chat:
                 attention_mask=attention_mask,
                 language_model_inputs=language_model_inputs,
                 generation_config=self.generation_config,
+                stopping_criteria=self.stopping_criteria,
                 return_dict_in_generate=True,
                 output_scores=True
             )