diff --git a/pytorch_forecasting/models/base/_base_model.py b/pytorch_forecasting/models/base/_base_model.py
index c4f30509f..8635a7c49 100644
--- a/pytorch_forecasting/models/base/_base_model.py
+++ b/pytorch_forecasting/models/base/_base_model.py
@@ -1755,7 +1755,15 @@ def predict(
         logging.getLogger("lightning").setLevel(log_level_lighting)
         logging.getLogger("pytorch_lightning").setLevel(log_level_pytorch_lightning)
 
-        return predict_callback.result
+        trainer_predict_callback = next(
+            (
+                callback
+                for callback in reversed(trainer.callbacks)
+                if isinstance(callback, PredictCallback)
+            ),
+            predict_callback,
+        )
+        return trainer_predict_callback.result
 
     def predict_dependency(
         self,
diff --git a/tests/test_models/test_temporal_fusion_transformer.py b/tests/test_models/test_temporal_fusion_transformer.py
index 8473dc2e7..3029e30d3 100644
--- a/tests/test_models/test_temporal_fusion_transformer.py
+++ b/tests/test_models/test_temporal_fusion_transformer.py
@@ -358,6 +358,19 @@ def test_prediction_with_dataloder(model, dataloaders_with_covariates, kwargs):
     model.predict(val_dataloader, fast_dev_run=True, **kwargs)
 
 
+def test_prediction_with_dataloader_twice(model, dataloaders_with_covariates):
+    val_dataloader = dataloaders_with_covariates["val"]
+    first_prediction = model.predict(
+        val_dataloader, fast_dev_run=True, return_index=True
+    )
+    second_prediction = model.predict(
+        val_dataloader, fast_dev_run=True, return_index=True
+    )
+
+    assert len(first_prediction.index) > 0
+    assert len(second_prediction.index) > 0
+
+
 def test_prediction_with_dataloder_raw(data_with_covariates, tmp_path):
     # tests correct concatenation of raw output
     test_data = data_with_covariates.copy()