reduce processing time

app.py

@@ -54,7 +54,7 @@ async def summarize(req: Request, body: SummarizeRequest):
     #     num_return_sequences=1
     # )
     outputs = model.generate(
-        input_ids=input_ids, attention_mask=attention_masks,
+        input_ids=input_ids, attention_mask=attention_mask,
         max_length=256,
         early_stopping=True
     )