Update handler.py

handler.py

@@ -15,8 +15,7 @@ class EndpointHandler():
         cfg = {
             "repo": "MrOvkill/Phi-3-Instruct-Bloated",
         }
-        self.model = AutoModelForCausalLM.from_pretrained(cfg['repo'], trust_remote_code=True, torch_dtype=torch.float16)
-        self.tokenizer = AutoTokenizer.from_pretrained(cfg['repo'])
+        self.pipe = pipeline("text-generation", "MrOvkill/Phi-3-Instruct-Bloated")
 
     def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
         """
@@ -38,10 +37,10 @@ class EndpointHandler():
             })
         
         res = PROMPT_FORMAT.format(inputs=data['inputs'])
-        return self.model(
+        return self.pipe(
             res,
-            max_new_tokens=max_new_tokens,
-            do_sample=False
+            do_sample=False,
+            max_new_tokens=max_new_tokens
         )
 
         return res