try torch compile

app.py

@@ -47,6 +47,13 @@ if torch.cuda.is_available():
     pipe.vae = AutoencoderTiny.from_pretrained(
         "madebyollin/taesd3", torch_dtype=torch.float16
     )
+
+    pipe.vae.decoder.layers = torch.compile(
+        pipe.vae.decoder.layers,
+        fullgraph=True,
+        dynamic=False,
+        mode="max-autotune-no-cudagraphs",
+    )
     pipe.vae.config.shift_factor = 0.0
 
     pipe = pipe.to(device)