Fix for fetching variants only

[DN6]

What does this PR do?

With PR: #9869 we fixed downloading sharded variants only and mixed variants, however we missed the case where a component might have sharded non-variant files and non-sharded variant file, which is the case with issue: #10634

This PR:

1. Adds a simpler has_variant check and removes the previous checks. This check first checks to see if we are in a component folder and then checks if any variants exist within the folder. If no folder exists then skip trying to add the additional non variant file.
2. Adds additional tests to check for the condition mentioned in issue The huggingface repo need to be fixed for Sana 2K and 4K models #10634.

Since usable_filenames is always populated with variants, we capture the necessary variant files and what we're trying to avoid is extra file downloads.

The only edge case I can think of here where this would fail (which passes with the current implementation) is if the filenames are the following:

Non-Variant in the main dir and a variant in a subfolder. Although I think this an edge case that we probably can't load anyway? I can't think of any pipelines that would have this configuration.

filenames = ["diffusion_pytorch_model.safetensors", f"unet/diffusion_pytorch_model.{variant}.safetensors"]

Fixes # (issue)
#10634

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.

[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.

[lawrence-cj]

Thanks @DN6 for supporting!

[yiyixuxu]

what happends if like we only have a bf16.bin and this is a non-variant safetensors?

[DN6]

As in we are trying to fetch something like this?

        variant = "fp16"
        filenames = [
                f"vae/diffusion_pytorch_model.{variant}.bin",
                f"text_encoder/model.{variant}.bin",
                f"unet/diffusion_pytorch_model.{variant}.bin",
        ]
        model_filenames, variant_filenames = variant_compatible_siblings(filenames, variant=None)

[yiyixuxu]

like this, I think we should fetch the non-variant safetensors in this case, no?

    variant = "fp16"
    filenames = [
            f"vae/diffusion_pytorch_model.{variant}.bin",
            f"text_encoder/model.{variant}.bin",
            f"unet/diffusion_pytorch_model.{variant}.bin",
            f"vae/diffusion_pytorch_model.safetensors",
            f"text_encoder/model.safetensors",
            f"unet/diffusion_pytorch_model.safetensors",
    ]

[DN6]

Hmm currently the behaviour on main is to return all the files in that list (both bin and safetensors) as usable_filenames and I think the ignore patterns would remove the bin files, resulting in just the safetensors being downloaded.

With this change only the fp16.bin files would be downloaded. Which feels technically "correct" to me since they are the "variant" files of each component. IMO non-variants should only be downloaded if no variant exists (regardless of format)

But this case implies that the proposal here is a breaking change, so I'll update to account for it.

[DN6]

This change is needed to correct a weird bug that was caught with this test
tests/pipelines/test_pipelines.py::CustomPipelineTests::test_load_custom_github

What was basically happening was that variant_compatible_siblings was incorrectly returning both the bin and safetensors version of the checkpoint in this repo
https://huggingface.co/google/ddpm-cifar10-32/tree/main

In the DiffusionPipeline download method, if model_folder_names ends up being an empty set

diffusers/src/diffusers/pipelines/pipeline_utils.py

Line 1414 in 33f9361

model_folder_names = {os.path.split(f)[0] for f in model_filenames if os.path.split(f)[0] in folder_names}

and is passed to is_safetensors_compatible the function will return False because filenames will end up being an empty set. But since variant_compatible_siblings was asking to fetch both bin and safetensors versions of the checkpoint, the test passed because the bin version was downloaded.

[DN6]

@yiyixuxu Had to do a bit more of a refactor to account for safetensors prioritization. But it should be much more robust to handle any number of repo file combinations. I've added a test for your case and a few others as well. I think they should cover all likely repo file layout scenarios, but if there are others I may have missed lmk.