diff --git a/src/transformers/models/perceiver/modeling_perceiver.py b/src/transformers/models/perceiver/modeling_perceiver.py
index 40d8f8fb9970..00ec22ba3fc3 100755
--- a/src/transformers/models/perceiver/modeling_perceiver.py
+++ b/src/transformers/models/perceiver/modeling_perceiver.py
@@ -1978,7 +1978,7 @@ def decoder_query(self, inputs, modality_sizes=None, inputs_without_pos=None, su
             # to get the indices for the unflattened array
             # unravel_index returns a tuple (x_idx, y_idx, ...)
             # stack to get the [n, d] tensor of coordinates
-            indices = torch.unravel_index(subsampled_points, self.output_index_dims)
+            indices = torch.unravel_index(subsampled_points.to(inputs.device), self.output_index_dims)
             pos = torch.stack(indices, dim=1)
             batch_size = inputs.shape[0]
             # Map these coordinates to [-1, 1]