try to figure out how ZeroGPU works

app.py

@@ -136,7 +136,7 @@ def read(prompt,
          emb1,
          emb2
          ):
-    with torch.inference_mode():
+    with torch.no_grad():
         sr, wav, fig = controllable_ui.read(prompt,
                                             language.split(" ")[-1].split("(")[1].split(")")[0],
                                             language.split(" ")[-1].split("(")[1].split(")")[0],