Fix Tiling in ConsistencyDecoderVAE

[tolgacangoz]

Fixes #6126.

This pull request includes three commits:

• Fix tiling

• Add docstring to decode method in ConsistencyDecoderVAE

I tested with this:

import torch
from diffusers import ConsistencyDecoderVAE

vae = ConsistencyDecoderVAE.from_pretrained("openai/consistency-decoder").to("cuda")
vae.enable_tiling()
vae.encode(torch.randn((1, 3, 256, 256), device="cuda"))

Also, this PR passes ./tests/models/autoencoders/test_models_vae.py tests in my local setup.

@yiyixuxu @sayakpaul

[sayakpaul]

This is looking solid.

I left a major comment on how we should expose the tiling arguments (init vs. via enable_vae_tiling()).

Can we also add a test for this?

[yiyixuxu]

hey! thanks for the PR.
it's looking great!

I left a feedback and additionally, could you add a test_vae_tiling test to ./tests/models/autoencoders/test_models_vae.py( similar to

diffusers/tests/pipelines/test_pipelines_common.py

Line 101 in d3986f1

def test_vae_tiling(self):

)

[sayakpaul]

Looking good! Let's add a test as well?

[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]

@StandardAI let us know when you are ready for a final review

[tolgacangoz]

It seems to pass the test in Colab.

[sayakpaul]

One comment and once addressed, we can ship!

[sayakpaul]

Please ping me once the CI run is done.

[tolgacangoz]

Please ping me once the CI run is done.

@sayakpaul

[sayakpaul]

Thank you for this impressive contribution!

[tolgacangoz]

Thanks so much for reviewing and merging!