Update handler.py

handler.py

@@ -8,8 +8,9 @@ api = HfApi(token=os.getenv("HUGGINGFACE_HUB_TOKEN"))
 
 class EndpointHandler():
     def __init__(self, path=""):
+        self.token = os.getenv("HUGGINGFACE_HUB_TOKEN")
         self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
-        self.model = LlamaForCausalLM.from_pretrained(path).to(self.device)
+        self.model = LlamaForCausalLM.from_pretrained(path,token=self.token).to(self.device)
         self.tokenizer = AutoTokenizer.from_pretrained(path)
 
     def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]: