use padding

interfaces/cap.py

@@ -105,7 +105,7 @@ def predict(text, model_id, tokenizer_id):
         text,
         max_length=256,
         truncation=True,
-        padding="do_not_pad",
+        padding=True,
         return_tensors="pt"
     )
     inputs = {k: v.to(device) for k, v in inputs.items()}