Update app.py

app.py

@@ -31,7 +31,7 @@ def load_model(model_name):
     model = AutoModelForCausalLM.from_pretrained(
         f"{root_path}/{model_name}",
         device_map="auto",
-        load_in_4bit=True,
+        # load_in_4bit=True,
         torch_dtype=torch.bfloat16,
         trust_remote_code=True,
     )