Fix compile in HunyuanVideo example

[eppaneamd]

In HunyuanVideo example, the model is compiled as follows:

        pipe.transformer = torch.compile(pipe.transformer,
                                         mode="max-autotune-no-cudagraphs")

However, when using compile and model cpu offloading together, this causes:

[rank6]: Traceback (most recent call last):
[rank6]:   File "/home/repos/xDiT/examples/hunyuan_video_usp_example.py", line 328, in <module>
[rank6]:     main()
[rank6]:   File "/home/repos/xDiT/examples/hunyuan_video_usp_example.py", line 277, in main
[rank6]:     output = pipe(
[rank6]:   File "/home/repos/pytorch/torch/utils/_contextlib.py", line 116, in decorate_context
[rank6]:     return func(*args, **kwargs)
[rank6]:   File "/usr/local/lib/python3.10/dist-packages/diffusers/pipelines/hunyuan_video/pipeline_hunyuan_video.py", line 682, in __call__
[rank6]:     self.maybe_free_model_hooks()
[rank6]:   File "/usr/local/lib/python3.10/dist-packages/diffusers/pipelines/pipeline_utils.py", line 1118, in maybe_free_model_hooks
[rank6]:     self.enable_model_cpu_offload(device=getattr(self, "_offload_device", "cuda"))
[rank6]:   File "/usr/local/lib/python3.10/dist-packages/diffusers/pipelines/pipeline_utils.py", line 1049, in enable_model_cpu_offload
[rank6]:     self.remove_all_hooks()
[rank6]:   File "/usr/local/lib/python3.10/dist-packages/diffusers/pipelines/pipeline_utils.py", line 1016, in remove_all_hooks
[rank6]:     accelerate.hooks.remove_hook_from_module(model, recurse=True)
[rank6]:   File "/home/repos/accelerate/src/accelerate/hooks.py", line 203, in remove_hook_from_module
[rank6]:     delattr(module, "_hf_hook")
[rank6]:   File "/home/repos/pytorch/torch/nn/modules/module.py", line 2040, in __delattr__
[rank6]:     super().__delattr__(name)
[rank6]: AttributeError: _hf_hook

Repro command:

torchrun --nproc_per_node=8 examples/hunyuan_video_usp_example.py --model tencent/HunyuanVideo \
--prompt "In the large cage, two puppies were wagging their tails at each other." \
--height 544 --width 960 --num_frames 129 --num_inference_steps 30 --warmup_steps 1 \
--ulysses_degree 8 --enable_tiling --enable_model_cpu_offload --use_torch_compile

This PR proposes modifying the model compilation to:

        pipe.transformer.compile()

Which does not cause the error. Also max-autotune is rather time-consuming process thus opting for default mode.

Additionally, there are currently some limitations for HunyuanVideo which are outlined in the README. They can be removed as soon as they have a fix implemented.

[StrongerXi]

Related: pytorch/pytorch#150711

If you upgrade pytorch, this should be fixed as well. But yes in general we recommend model.compile over torch.compile(model).