Enable BitsAndBytes quantization in MPS

[LucasSte]

What does this PR do?

Bitsandbytes now has basic support for the Apple MPS backend, as I can tell by bitsandbytes-foundation/bitsandbytes#1818 and
bitsandbytes-foundation/bitsandbytes#1875.

The issue is that diffusers does not allow me to use the quantization on Apple hardware, because of the error No GPU found. A GPU is needed for quantization. from here

diffusers/src/diffusers/quantizers/bitsandbytes/bnb_quantizer.py

Lines 64 to 65 in f2be8bd

	if not (torch.cuda.is_available() or torch.xpu.is_available()):
	raise RuntimeError("No GPU found. A GPU is needed for quantization.")

.

Adding MPS to that check, in addition to the other suggested change in #13361 (comment), allows us to enable BitsAndBytes for MPS.

I tested the change with the quantized version of FLUX.2-dev as described in https://github.com/black-forest-labs/flux2/blob/main/docs/flux2_dev_hf.md#4-bit-transformer-and-4-bit-text-encoder-20g-of-vram, and all worked fine.

This PR fixes #13361.

Before submitting

• [X ] This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
  => Not the case.
• [X ] Did you read the contributor guideline?
• [X ] Did you read our philosophy doc (important for complex PRs)?
• [X ] Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
  => Enable MPS backend for bitsandbytes quantization #13361
• [X ] Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
  => I haven't found any documentation that mentions supported backends on BitsAndBytes in the diffusers repository, so I didn't modify anything in the docs folder.
• [X ] Did you write any new necessary tests?
  => I didn't write any tests, because the changes were trivial, but I can write them if reviewers deem necessary.

Who can review?

@sayakpaul
@asomoza

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[sayakpaul]

Thanks! Could we also see some results from MPS? @asomoza would you be able to quickly do a test?

[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.

[bghira]

jfyi these BNB kernels are worse than just using torchao or quanto int8 because they're very lacking in optimisation

[iwr-redmond]

There was an alternative offered that could have been more performant (bitsandbytes#1853), but in either case this PR is still needed to enable general use and future optimization.

[bghira]

for UX it should imo be converted to a warning instead of an exception, that the kernels are slow and experimental

[matthewdouglas]

FWIW, there will be some improvement soon to MPS backend in our next bitsandbytes release. We've merged these so far since the last release:

• Simplest MPS backend: adding kernels using kernel builder and kernels-community bitsandbytes-foundation/bitsandbytes#1875
• perf: improve default/fallback backend implementation for blockwise quantization ops bitsandbytes-foundation/bitsandbytes#1960

With kernels available, on macOS 26+ we'll pull down some kernels hosted on the HF Hub. These are mostly focused on 4bit.

Separately, we'll continue to build on PR 1960 with a few more improvements, for the torch.compile fallback implementations which will be used on macOS 14/15, and when Hub kernels cannot be loaded.

So, I think it is reasonable to enable it. If anything, maybe guard on bitsandbytes >= 0.49.0 as the absolute floor. It's not clear to me if it's the responsibility of diffusers to warn you if it is going to be slow or not.

[sayakpaul]

@Vargol thanks for your post but I had to delete because it looks NSFW to my eyes. Just letting you know for transparency.

[sayakpaul]

@asomoza a gentle reminder.

[bghira]

damn, requiring macos 26? but why? that causes a lockout of async simd as Apple has not provided any replacement.

[asomoza]

@sayakpaul sorry I missed this, sadly currently I don't have any method to test MPS.

[sayakpaul]

You mean any device? But anyway the changes in the PR look good enough to me to unblock MPS. So, will merge,