could not do any cuda operators when try to schedule with dp attention

[Hanatan1th]

I can run successfully with tp2 dp1 and enable_dp_attention=False, when I try to run with enable_dp_attention=True, tp4, dp2, will report like

when I change this torch.zeros to torch.empty, it will not report error here, but report similar error later

or

it seems I can only do cuda operators inside model forward when run with enable_dp_attention