use auth enabled

src/app.py

@@ -114,8 +114,10 @@ sample_data = [
 model_id = "ibm-granite/granite-docling-258M"
 
 if gr.NO_RELOAD:
-    processor = AutoProcessor.from_pretrained(model_id)
-    model = Idefics3ForConditionalGeneration.from_pretrained(model_id, device_map=device, torch_dtype=torch.bfloat16)
+    processor = AutoProcessor.from_pretrained(model_id, use_auth_token=True)
+    model = Idefics3ForConditionalGeneration.from_pretrained(
+        model_id, device_map=device, torch_dtype=torch.bfloat16, use_auth_token=True
+    )
     if not torch.cuda.is_available():
         model = model.to(device)