Update app.py

app.py

@@ -30,7 +30,11 @@ def _eos_ids(tok):
         ids.add(im_end)
     return list(ids)
 
-@spaces.GPU(duration=120)  # REQUIRED for ZeroGPU; remove if using standard GPU hardware
+@spaces.GPU()  # REQUIRED for ZeroGPU; remove if using standard GPU hardware
+def gradio_fn(message, history):
+    response = infer_text(history + [(message, None)])
+    return response
+
 def chat_fn(message, history, system_text, temperature, top_p, max_new, min_new):
     msgs = _msgs_from_history(history, system_text) + [{"role": "user", "content": message}]
     prompt = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)