Upload handler.py

handler.py

@@ -18,7 +18,7 @@ class EndpointHandler():
                     device_map="cuda",
                     trust_remote_code=True,
                     fp16=True).eval()
-        tokenizer = transformers.AutoTokenizer.from_pretrained(
+        self.tokenizer = transformers.AutoTokenizer.from_pretrained(
                     path,
                     cache_dir=None,
                     model_max_length=2048,
@@ -26,7 +26,7 @@ class EndpointHandler():
                     use_fast=False,
                     trust_remote_code=True,
                 )
-        tokenizer.pad_token_id = tokenizer.eod_id
+        self.tokenizer.pad_token_id = self.tokenizer.eod_id
         return
  
     def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]: