app.py correct limit hf

app.py

@@ -81,6 +81,11 @@ def convert():
 DOC_DB_LIMIT = 10
 MAX_NEW_TOKENS = 2048
 
+# Limit in HF, no need to set it
+if "SET_LIMIT" == os.getenv("DEMO"):
+    DOC_DB_LIMIT = 4
+    MAX_NEW_TOKENS = 32
+
 with gr.Blocks(theme=theme, css=css) as demo:
   with gr.Tab("Chat"):
 
@@ -122,10 +127,5 @@ with gr.Blocks(theme=theme, css=css) as demo:
   change_model_button.click(dc.change_llm,[repo_, file_, max_tokens, temperature, top_p, top_k, repeat_penalty, max_docs],[model_verify])
 
   falcon_button.click(dc.default_falcon_model, [], [model_verify])
-
-# limit in HF, no need to set it
-if "SET_LIMIT" == os.getenv("DEMO"):
-    DOC_DB_LIMIT = 4
-    MAX_NEW_TOKENS = 32
     
 demo.launch(enable_queue=True)