use local model

app.py

@@ -17,7 +17,8 @@ As a derivate work of [ConsistentAgents]() by Seonghee Lee.
 
 if torch.cuda.is_available():
     model_id = "./backprop_llama2_69_1e-05"
-    model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
+    HF_ACCESS_TOKEN = os.getenv('HF_ACCESS_TOKEN')
+    model = AutoModelForCausalLM.from_pretrained(model_id, use_auth_token=HF_ACCESS_TOKEN, torch_dtype=torch.float16, device_map="auto")
     tokenizer = AutoTokenizer.from_pretrained(model_id)
     tokenizer.use_default_system_prompt = False