try to figure out how ZeroGPU works

app.py

@@ -117,6 +117,9 @@ class ControllableInterface(torch.nn.Module):
                                   loudness_in_db=loudness_in_db)
         return sr, wav, fig
 
+@spaces.GPU
+def get_gw():
+    return GanWrapper(os.path.join(MODELS_DIR, "Embedding", "embedding_gan.pt"), device="cuda" if torch.cuda.is_available() else "cpu")
 
 title = "Controllable Text-to-Speech for over 7000 Languages"
 article = "Check out the IMS Toucan TTS Toolkit at https://github.com/DigitalPhonetics/IMS-Toucan"
@@ -124,7 +127,7 @@ available_artificial_voices = 1000
 path_to_iso_list = "Preprocessing/multilinguality/iso_to_fullname.json"
 iso_to_name = load_json_from_path(path_to_iso_list)
 text_selection = [f"{iso_to_name[iso_code]} Text ({iso_code})" for iso_code in iso_to_name]
-gw = GanWrapper(os.path.join(MODELS_DIR, "Embedding", "embedding_gan.pt"), device="cuda" if torch.cuda.is_available() else "cpu")
+gw = get_gw()
 controllable_ui = ControllableInterface(gan_wrapper=gw, available_artificial_voices=available_artificial_voices)