Update app.py

app.py

@@ -56,13 +56,13 @@ print(f"  device={device}")
 
 model_ledger = ModelLedger(
     dtype=torch.bfloat16,
-    device=device,
+    device="cpu",
     checkpoint_path=checkpoint_path,
     gemma_root_path=gemma_local_path,
 )
 
 # Load text encoder once and keep it in memory
-text_encoder = model_ledger.text_encoder()
+text_encoder = model_ledger.text_encoder().to("cuda")
 
 print("=" * 80)
 print("Text encoder loaded and ready!")