I suggest that you file a bug. The bug report could be fairly simple, linking back to this post if you wish.
I see the same behavior on CUDA 13, L4 GPU. I note that if I run your test case with compute-sanitizer I can get additional information about what may be going wrong.