[AttentionMaskConverter] fix sdpa unmask unattended (#28369)

fix tensor device

src/transformers/modeling_attn_mask_utils.py

@@ -234,8 +234,8 @@ class AttentionMaskConverter:
 
         # Get the index of the first non-zero value for every sample in the batch.
         # In the above example, indices = [[2], [0], [1]]]
-        tmp = torch.arange(attention_mask.shape[1], 0, -1)
-        indices = torch.argmax(attention_mask.cpu() * tmp, 1, keepdim=True)
+        tmp = torch.arange(attention_mask.shape[1], 0, -1, device=attention_mask.device)
+        indices = torch.argmax(attention_mask * tmp, 1, keepdim=True)
 
         # Find the batch indexes that have unattended tokens on the leftmost side (e.g. [0, 0, 1, 1, 1]), for which the first rows of the
         # expanded mask will be completely unattended.