Update app.py

app.py

@@ -85,6 +85,7 @@ def loadModel():
         #        )
         tokenizer =  AutoTokenizer.from_pretrained(base_model,use_fast=False,cache_dir=cache_dir)
         tokenizer.pad_token = tokenizer.unk_token
+        model.to('cuda')
     return model, tokenizer
 
 model, tokenizer = loadModel()