edit

api.py

@@ -50,7 +50,9 @@ async def create_item(request: Request):
 
 
 if __name__ == '__main__':
-    model_file = "chatglm-6b"
+    #model_file = "chatglm-6b"
+    #model_file = "THUDM/chatglm-6b-int4-qe"
+    model_file = "chatglm-6b-int4-qe"
     tokenizer = AutoTokenizer.from_pretrained(model_file, trust_remote_code=True)
     #quantization_config= BitsAndBytesConfig(load_in_8bit=True)