Update handler.py

handler.py

@@ -14,6 +14,6 @@ class EndpointHandler:
         for q in inp:
             inp[q] = inp[q].to(device)
         with torch.inference_mode():
-            out= model.generate(**inp)
-        final_output = tokenizer.batch_decode(out,skip_special_tokens=True)
+            out= self.model.generate(**inp)
+        final_output = self.tokenizer.batch_decode(out,skip_special_tokens=True)
         return {"translation": final_output[0]}