Kandinsky 5.0 Video Pro and Image Lite

[leffff]

This PR:

1. Updates Kandinsky 5.0 Video code for handling Pro Model.
2. Adds Kandinsky 5.0 Image Lite (T2I and I2I)
   @yiyixuxu @sayakpaul

[sayakpaul]

pipeline = pipeline.to("cuda:0")

We actually don't need it if we're doing model CPU offloading. Additionally, we can do compile_repeated_blocks() instead of compile(). This will reduce the compilation time substantially!

Could you try that?

[leffff]

Right now there is some issue with compile_repeated_blocks()
I will look into that on Monday.

[leffff]

pipeline = pipeline.to("cuda:0")

We actually don't need it if we're doing model CPU offloading. Additionally, we can do compile_repeated_blocks() instead of compile(). This will reduce the compilation time substantially!

Could you try that?

I guess, we need it to specify the device. In case of a multi GPU machine. Can we leave it as is?

[sayakpaul]

.to() won't be needed in any case. enable_model_cpu_offload() already defaults to the 0th ID and "cuda" as its device:

https://github.com/huggingface/diffusers/blob/544ba677dd97a49c8124208837025aa8b5ab639e/src/diffusers/pipelines/pipeline_utils.py#L1150C9-L1150C33

Otherwise, if we want to be more explicit, then we can add the following at the beginning of the script:

import os
os.environ["CUDA_VISIBLE_DEVICES"] = str(0)

[leffff]

Great, we have resolved the problems

[yiyixuxu]

@bot /style

[github-actions]

Style fix is beginning .... View the workflow run here.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[yiyixuxu]

can you fix the kensindky tests? you can test locally with

pytest tests/pipelines/kandinsky5/test_kandinsky5.py

[leffff]

can you fix the kensindky tests? you can test locally with

pytest tests/pipelines/kandinsky5/test_kandinsky5.py

Sure