Update app.py

app.py

@@ -304,7 +304,7 @@ sampling_params = {
 }
 
 # Initialize LLM
-llm = LLM(max_model_len=8096)
+llm = LLM(max_model_len=32000)
 
 with gr.Blocks() as demo:
     gr.Markdown(f"<h2>{NAME_OF_SERVICE}</h2>")