Update app.py

app.py

@@ -60,7 +60,7 @@ endpoint_url = os.environ.get("ENDPOINT_URL", "http://127.0.0.1:8080")
 client = Client(endpoint_url, timeout=120)
 eos_token = "</s>"
 MAX_MAX_NEW_TOKENS = 1024
-DEFAULT_MAX_NEW_TOKENS = 256
+DEFAULT_MAX_NEW_TOKENS = 1024
 
 max_prompt_length = 4096 - MAX_MAX_NEW_TOKENS - 10
 
@@ -237,5 +237,5 @@ with gr.Blocks() as demo:
 
     gr.Markdown(LICENSE)
 
-demo.queue(max_size=128)
+demo.queue(concurrency_count=4, max_size=128)
 demo.launch()