Update handler.py

handler.py

@@ -13,10 +13,12 @@ os.environ["CUDA_VISIBLE_DEVICES"] = "0"
 class EndpointHandler:
     def __init__(self, path=""):
         model_bin_path = os.path.join(path, "model.bin")
+        model_folder_path = model_name = os.path.join(path, "Wizard-Vicuna-13B-Uncensored-GPTQ")
+        
         with open(model_bin_path, "rb") as f: # "rb" because we want to read in binary mode
             self.model = pickle.load(f)
 
-        self.tokenizer = AutoTokenizer.from_pretrained("Wizard-Vicuna-13B-Uncensored-GPTQ", use_fast=False)
+        self.tokenizer = AutoTokenizer.from_pretrained(model_folder_path, use_fast=False)