diff --git a/tests/test_ops/test_iou3d.py b/tests/test_ops/test_iou3d.py
index 37fa417406b..6b4d552e8cf 100644
--- a/tests/test_ops/test_iou3d.py
+++ b/tests/test_ops/test_iou3d.py
@@ -96,14 +96,18 @@ def test_nms3d(device):
     assert np.allclose(inds.cpu().numpy(), np_inds)
 
     # test for many boxes
-    # np.random.seed(42)
-    # np_boxes = np.random.rand(555, 7).astype(np.float32)
-    # np_scores = np.random.rand(555).astype(np.float32)
-    # boxes = torch.from_numpy(np_boxes)
-    # scores = torch.from_numpy(np_scores)
-    # inds = nms3d(boxes.to(device), scores.to(device), iou_threshold=0.3)
-
-    # assert len(inds.cpu().numpy()) == 176
+    # In the float data type calculation process, sometimes float will be
+    # converted to double in CUDA kernel, always use float in MLU kernel.
+    # The difference between the mentioned above leads to different results.
+    if device != 'mlu':
+        np.random.seed(42)
+        np_boxes = np.random.rand(555, 7).astype(np.float32)
+        np_scores = np.random.rand(555).astype(np.float32)
+        boxes = torch.from_numpy(np_boxes)
+        scores = torch.from_numpy(np_scores)
+        inds = nms3d(boxes.to(device), scores.to(device), iou_threshold=0.3)
+
+        assert len(inds.cpu().numpy()) == 176
 
 
 @pytest.mark.parametrize('device', [