Contribute working version of Stable Diffusion on Mac OSX M3 Max

[chrismattmann]

This PR contributes a working version of Stable Diffusion notebook for Mac OS X M3 Max.

Fixes #9967 Tried to run diffusers/stable_diffusion.ipynb on Mac M3 and ran into errors

Who can review?

@pcuenca @sayakpaul

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[pcuenca]

Thank you @chrismattmann! Did you have to make any compatibility changes, other than changing the compute device to mps (and ignoring nvidia-smi)? If so, perhaps it could be more interesting to bring the device selection (cuda or mps) to the original notebook so any future updates are easier to maintain.

[chrismattmann]

@pcuenca good point, I think that was the only changes was device selection. I suppose I could parameterize it. I will take a look at doing that this weekend and send an update!

[chrismattmann]

OK @pcuenca @sayakpaul updated in a52e0e5 please review

[sayakpaul]

I will let @pcuenca to review and merge :)

[chrismattmann]

I will let @pcuenca to review and merge :)

Thank you! @pcuenca can you please take a look? Thanks @sayakpaul !

[sayakpaul]

Please allow till this week to review. Feel free to re-ping after that :-)

[chrismattmann]

@pcuenca @sayakpaul sorry to bother you but you mentioned to re-ping in a week, so I'm doing that. Thanks!

[pcuenca]

Hello @chrismattmann! Thanks for the ping, and sorry for being slow to review 🙏

The PR is looking good, I just have a few minor comments:

• Instead of using platform, I would show the use of nvidia-smi if this condition succeeds:

if torch.cuda.is_available():
    # Shows the nVidia GPUs, if this system has any
    !nvidia-smi

• Similarly, I would verify that the mps device is indeed available. Note that the platform being Darwin does not guarantee the existence of such device. So I would use something like this:

    if torch.cuda.is_available():
        device=torch.device("cuda")
    elif torch.backends.mps.is_available():
        device=torch.device("mps")

With this change, you no longer need the isosx variable and can move tensors and pipelines .to(device) everywhere.

• You can also use the device variable, as used above, instead of defining another one (generator_type) for the generator. The following should work:

generator = torch.Generator(device).manual_seed(1024)

Thanks a lot for your patience, I believe we are very close to finally merging this!

[chrismattmann]

thanks @pcuenca I will make these updates and contribute an update this week!

[pcuenca]

Thanks a lot @chrismattmann, feel free to ping me when you do 🙌

[chrismattmann]

Dear @pcuenca I've gone ahead and addressed the comments. Please review and merge! :)

[pcuenca]

Thanks a lot, @chrismattmann! 🙌 🔥

[chrismattmann]

Thank you @pcuenca ! appreciate you!

[chrismattmann]

Thank you and this closes #9967