Dataloading hangs on Trillium when using recent whls

[miladm]

Dataloading hangs on Trillium when using recent torch and torchxla whls.

Passing to @bhavya01 to offer detailed context.

[hsebik]

Hi Milad,
I have similar experience on Trillium not in dataloading, it is on diffusers inference. Parallel spawned processes are going zombie after couple of inference. Do you have any observation on your side?