diff --git a/fastvideo/eval/metrics/vbench/scene/metric.py b/fastvideo/eval/metrics/vbench/scene/metric.py
index d8dc70354a..77c3fe4abb 100644
--- a/fastvideo/eval/metrics/vbench/scene/metric.py
+++ b/fastvideo/eval/metrics/vbench/scene/metric.py
@@ -63,7 +63,7 @@ def setup(self) -> None:
         self._model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
             self._model_path,
             torch_dtype=torch.bfloat16,
-            device_map=str(self.device) if self.device.type == "cuda" else None,
+            device_map=str(self.device) if self.device.type != "cpu" else None,
         )
         self._model.disable_talker()
         self._model.eval()