Added cpu offload to enable full length 720p on a 4090.

[XmYx]

[MikeAiJF]

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument tensors in method wrapper_CUDA_cat)

[MikeAiJF]

[MikeAiJF]

大神，这个应该怎么解决

[FrankLeeeee]

I would recommend that we use individual inference proposed in https://github.com/hpcaitech/Open-Sora/pull/597 instead of interrupting the scheduler's internal forward.

[zhengzangw]

Closed as we updated to Open Sora 2.0