Could I use the device map for pipelines parallel training?

[enze]

Is this feature used for pipeline parallel training？