[Single File] Add GGUF support

[DN6]

What does this PR do?

Adds support for loading GGUF checkpoints via from_single_file.

import torch

from diffusers import FluxPipeline, FluxTransformer2DModel, GGUFQuantizationConfig

ckpt_path = (
    "https://huggingface.co/city96/FLUX.1-dev-gguf/blob/main/flux1-dev-Q2_K.gguf"
)
transformer = FluxTransformer2DModel.from_single_file(
    ckpt_path,
    quantization_config=GGUFQuantizationConfig(compute_dtype=torch.bfloat16),
    torch_dtype=torch.bfloat16,
)
pipe = FluxPipeline.from_pretrained(
    "black-forest-labs/FLUX.1-dev",
    transformer=transformer,
    generator=torch.manual_seed(0),
    torch_dtype=torch.bfloat16,
)
pipe.enable_model_cpu_offload()
prompt = "A cat holding a sign that says hello world"
image = pipe(
    prompt_embeds=prompt_embeds, pooled_prompt_embeds=pooled_prompt_embeds
).images[0]
image.save("flux-gguf.png")

Notes:

1. API for loading GGUF is a bit overkill, but it's consistent with quantized loading in from_pretrained. GGUF files have enough metadata that we can automatically infer everything we need from the file itself. We don't really need a quantization config, but it becomes necessary as we expand to support to other quant loading methods (BnB, TorchAO etc)

TODOS:

• Benchmark loading and inference speed.
• Verify output quality
• Add tests

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

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.

[DN6]

GGUF needs to access the tensor quant type to run a shape check. So this needs to change from passing in shapes to passing in params directly.

[sayakpaul]

Why not add this method to the gguf_quantizer.py file instead of modifying this? This would be a breaking change no?

[sayakpaul]

I see you're already adding this to the GGUF quantizer class. So, maybe okay to not modify this?

[DN6]

GGUF files contain a mix of quantized linear and unquantized linear layers. It's not trivial to selectively replace layers. We can replace all of them and then check the parameter type when running forward instead.

[DN6]

nvm

[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.

[sayakpaul]

Suggested change

	# Force requires grad to False to avoid unexpected errors
	# Force requires_grad to False to avoid unexpected errors