update app.py

app.py

@@ -17,7 +17,8 @@ def load_model(model_path):
     model = AutoModelForCausalLM.from_pretrained(
         config.base_model_name_or_path,
         quantization_config=bnb_config,
-        trust_remote_code=True
+        trust_remote_code=True,
+        token=os.environ["HF_TOKEN"],
     )
     model.config.use_cache = False
     tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path)