Token Issue with FluxTransformer2DModel.from_single_file Method

[CyberVy]

Describe the bug

When I use the FluxTransformer2DModel.from_single_file method, the provided token seems to be invalid, and I am unable to download files that require token verification, such as the official Flux model files.

However, I can confirm that the token is valid because when using FluxPipeline.from_single_file to load the official Flux model, all files can be downloaded successfully without any issues.

Reproduction

from diffusers import FluxTransformer2DModel
FluxTransformer2DModel.from_single_file(“A Local File Path”,token=...)

It does not work, and raise an error in the log below.

from diffusers import FluxTransformer2DModel
FluxTransformer2DModel.from_single_file(“A HF URL”,token=...)

It works too.
So I think the problem only exists when a local file path is passed to this method.
When a HF url is passed to it, it works.

from diffusers import FluxPipeline
FluxPipeline.from_single_file("black-forest-labs/FLUX.1-dev",token=...)

It works.

Logs

...
HTTPError: 401 Client Error: Unauthorized for url: https://huggingface.co/black-forest-labs/FLUX.1-dev/resolve/main/transformer/config.json
...
OSError: black-forest-labs/FLUX.1-dev is not a local folder and is not a valid model identifier listed on 'https://huggingface.co/models'
If this is a private repository, make sure to pass a token having permission to this repo with `token` or log in with `huggingface-cli login`.

System Info

• 🤗 Diffusers version: 0.31.0
• Platform: Linux-6.6.56+-x86_64-with-glibc2.35
• Running on Google Colab?: No
• Python version: 3.10.14
• PyTorch version (GPU?): 2.4.0 (True)
• Flax version (CPU?/GPU?/TPU?): 0.8.4 (gpu)
• Jax version: 0.4.26
• JaxLib version: 0.4.26.dev20240620
• Huggingface_hub version: 0.25.1
• Transformers version: 4.45.1
• Accelerate version: 0.34.2
• PEFT version: 0.13.2
• Bitsandbytes version: not installed
• Safetensors version: 0.4.5
• xFormers version: not installed
• Accelerator: Tesla T4, 15360 MiB
• Tesla T4, 15360 MiB
• Using GPU in script?: No
• Using distributed or parallel set-up in script?: No

Who can help?

@DN6 @sayakpaul

[asomoza]

I tried with this code:

from diffusers import FluxTransformer2DModel


FluxTransformer2DModel.from_single_file(
    "https://huggingface.co/black-forest-labs/FLUX.1-dev/flux1-dev.safetensors",
    token="hf_XXX",
)

With a valid token and it worked without a problem, I can only guess that your problem is still the problem but it's weird that it works with the pipeline and not the transformer model.

You didn't post the diffusers version, are you using the latest one?

[sayakpaul]

it's weird that it works with the pipeline and not the transformer model.

In your code example, you used the model class, though and you said it worked?

[asomoza]

yeah, I just loaded it like OP was doing it, didn't do anything else since the issue was supposed to be the access token, is that your question?

[CyberVy]

I tried with this code:

from diffusers import FluxTransformer2DModel


FluxTransformer2DModel.from_single_file(
    "https://huggingface.co/black-forest-labs/FLUX.1-dev/flux1-dev.safetensors",
    token="hf_XXX",
)

With a valid token and it worked without a problem, I can only guess that your problem is still the problem but it's weird that it works with the pipeline and not the transformer model.

You didn't post the diffusers version, are you using the latest one?

Yes, it's the latest 0.31.0.
Copied from Kaggle console, there is some bug to copy information from kaggle console, so I missed some information.
Sorry for that.
Here is the complete env information.

• 🤗 Diffusers version: 0.31.0
• Platform: Linux-6.6.56+-x86_64-with-glibc2.35
• Running on Google Colab?: No
• Python version: 3.10.14
• PyTorch version (GPU?): 2.4.0 (True)
• Flax version (CPU?/GPU?/TPU?): 0.8.4 (gpu)
• Jax version: 0.4.26
• JaxLib version: 0.4.26.dev20240620
• Huggingface_hub version: 0.25.1
• Transformers version: 4.45.1
• Accelerate version: 0.34.2
• PEFT version: 0.13.2
• Bitsandbytes version: not installed
• Safetensors version: 0.4.5
• xFormers version: not installed
• Accelerator: Tesla T4, 15360 MiB
  Tesla T4, 15360 MiB
• Using GPU in script?: No
• Using distributed or parallel set-up in script?: No

[CyberVy]

I tried with this code:

from diffusers import FluxTransformer2DModel


FluxTransformer2DModel.from_single_file(
    "https://huggingface.co/black-forest-labs/FLUX.1-dev/flux1-dev.safetensors",
    token="hf_XXX",
)

With a valid token and it worked without a problem, I can only guess that your problem is still the problem but it's weird that it works with the pipeline and not the transformer model.

You didn't post the diffusers version, are you using the latest one?

Also, I've tried again, you code works.
The problem only exists when a local file path is passed to FluxTransformer2DModel.from_single_file method. When a HF url is passed to it, it works.
And I updated the issue with more detailed triggering conditions in the reproduction code.
Thanks for your help.

[CyberVy]

AFAIK
The problem only exists when a local file path is passed to FluxTransformer2DModel.from_single_file method. When a HF url is passed to it, it works.
I believe the issue essentially lies in the diffusers.loaders.single_file_model.FromOriginalModelMixin.from_single_file method.
Moreover, it doesn't seem that all from_single_file methods are affected, as so far, when I use diffusers.loaders.single_file.FromSingleFileMixin.from_single_file to load pipelines, no issues have occurred.
@sayakpaul

[DN6]

Thanks for flagging @CyberVy. Opened a PR to fix
#10082

When we try to auto fetch the configs needed to load the Flux model we weren't passing in the token. My guess is that you don't have the Flux configs in your HF Cache.

You can work around it by either setting the HF_TOKEN enviroment variable with your token or running huggingface-cli login and setting your token that way.

[yiyixuxu]

closing as fixed here #10082
let us know if you still have issues and we will reopen!