diff --git a/modelscope/pipelines/cv/ocr_recognition_pipeline.py b/modelscope/pipelines/cv/ocr_recognition_pipeline.py
index 4b095042..c20d020c 100644
--- a/modelscope/pipelines/cv/ocr_recognition_pipeline.py
+++ b/modelscope/pipelines/cv/ocr_recognition_pipeline.py
@@ -91,7 +91,8 @@ class OCRRecognitionPipeline(Pipeline):
                 data.append(mask)
 
         data = torch.FloatTensor(data).view(
-            len(data), 1, IMG_HEIGHT, IMG_WIDTH).cuda() / 255.
+            len(data), 1, IMG_HEIGHT, IMG_WIDTH) / 255.
+        data = data.to(self.device)
 
         result = {'img': data}