add GPU

app.py

@@ -25,6 +25,7 @@ faiss_index_path, chunks_path = prepare_index_and_chunks(
 model, tokenizer = load_model(MODEL_NAME)
 
 # ——— INFERENCE FUNCTION ———
+@spaces.GPU()
 def answer_query(query: str) -> str:
     if not query.strip():
         return "⚠️ Please enter a question."
@@ -51,4 +52,4 @@ iface = gr.Interface(
 )
 
 if __name__ == "__main__":
-    iface.launch()
+    iface.launch(share=True)