update

app.py

@@ -38,7 +38,7 @@ from litgpt.generate.base import sample
 
 device = "cuda" if torch.cuda.is_available() else "cpu"
 ckpt_dir = "./checkpoint"
-streaming_output = False
+streaming_output = True
 
 
 OUT_CHUNK = 4096
@@ -335,5 +335,7 @@ demo = gr.Interface(
     # live=True,
 )
 
-demo.queue()
-demo.launch()
+
+if __name__ == '__main__':
+    demo.queue()
+    demo.launch()