demo buildout 10

app.py

@@ -53,7 +53,7 @@ def infer(
     with torch.inference_mode():
         generated_ids = model.generate(
             **inputs,
-            max_new_tokens=max_new_tokens,
+            max_length=2048
         )
     result = processor.batch_decode(generated_ids, skip_special_tokens=True)
     return result[0][len(text):].lstrip("\n")