diffusers.FluxPipeline has no .from_single_file() function

[lutusp]

Describe the bug

At the moment one cannot use a local Flux1-dev or -schnell file for lack of an essential function "from_single_file()" in the diffusers.FluxPipeline library. The provided flag "local_files_only" for some reason doesn't solve this problem.

Reproduction

pipe = FluxPipeline.from_single_file(
 'full path to local file, example /(full_path)/flux1-dev.safetensors',
 torch_dtype=torch.bfloat16,
 use_safetensors=True,
 local_files_only=True
 )

Logs

No response

System Info

$ diffusers-cli env

• 🤗 Diffusers version: 0.30.0.dev0
• Platform: Linux-5.15.0-113-generic-x86_64-with-glibc2.35
• Running on Google Colab?: No
• Python version: 3.10.12
• PyTorch version (GPU?): 2.2.2+cu121 (True)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Huggingface_hub version: 0.24.5
• Transformers version: 4.41.1
• Accelerate version: 0.30.1
• PEFT version: not installed
• Bitsandbytes version: not installed
• Safetensors version: 0.4.3
• xFormers version: 0.0.25.post1
• Accelerator: NVIDIA GeForce RTX 4090, 24564 MiB
• Using GPU in script?: YES
• Using distributed or parallel set-up in script?: NO

Who can help?

No response

[DN6]

@lutusp Are you trying to load just the transformer model or do you have the weights for the entire checkpoint saved in a single file checkpoint?

[lutusp]

I guess my explanation wasn't clear enough. My point wasn't that the function created an unexpected outcome, it is that the function doesn't exist. There is no function with that name in diffusers.FluxPipeline.

But apart from that, it was this file: "flux1-dev.safetensors", from here: https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main .

[mlick]

I have a same question, i want to know how to load local model?

[DN6]

@lutusp Support for single file for Flux at the pipeline level hasn't been added yet because checkpoints that contain all the model components in a single file were only recently released. The checkpoint you shared contains just the transformer weights.

Support for single file loading just the Transformer exists currently.

transformer = FluxTransformer2DModel.from_single_file(
 "https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/flux1-dev.safetensors",
 torch_dtype=torch.bfloat16
).to("cuda")
pipe = FluxPipeline.from_pretrained(
 "black-forest-labs/FLUX.1-dev",
 transformer=transformer,
 torch_dtype=torch.bfloat16
)

[tomsin9]

for the FluxTransformer2DModel.from_single_file(),
when i change the path to local, it doesnt work:

OSError: runwayml/stable-diffusion-v1-5 does not appear to have a file named config.json.

[sxy771]

for the FluxTransformer2DModel.from_single_file(), when i change the path to local, it doesnt work:

OSError: runwayml/stable-diffusion-v1-5 does not appear to have a file named config.json.

I'm also wondering how to handle above errors.

I tried to load flux model downloaded from civitai but the model didn't work showing OSError message.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. If you think this still needs to be addressed please comment on this thread.

Please note that issues that do not follow the contributing guidelines are likely to be ignored.

[missionfloyd]

This isn't stale. Flux, like various other pipelines, still lacks single file support.

Most models seem to be distributed as safetensors files, so I think that the ability to use them should be a higher priority, and that diffusers shouldn't be so helplessly dependent on huggingface.co.

[asomoza]

Hi, you're mixing a lot of stuff.

The first and most important answer is that Flux has from_single_file support, it was added with this PR and this one too, this was added over a month ago so I really don't understand why are you posting this now.

Also to help understanding some concepts if you want to read after the answer.

safetensors is a format for saving the model weights, diffusers uses this format. Some old models and researches don't use the safetensors format and instead use the pickle format.

Diffusers is not "helplessly dependent on huggingface.co" as you mention, we promote the use of it because it makes things a lot easier for everyone that works with a lot of models and doesn't have the time to track and download each model from different websites and also it helps testing and solving issues by a lot since we don't have to manually download what the users are using when they have problems. It also helps people that know how to use git to update with just a single 2 word command line instead of going to some random site and check if the model was updated and then manually download each version.

You can use from_pretrained with local multi directory diffusers models or use from_single_file with also local models, you don't have to download them from the huggingface hub.

We stated this before and I repeat it here, we are giving the same importance to single file support as the multi directory diffusers format. The only reason why Flux didn't have single file support was because the original release wasn't with a single file, BFL released just the transformer model and the AE, people needed to download the T5 separately to use it.

Later some people started using the single file format and not everyone uses it as you say, because it's a big waste of space to have a 10GB text encoder for every model when all of them use the same one.

[asomoza]

I removed the stale tag just in case if you have some problem with the from_single_file, but if there isn't a problem anymore, I will close this issue as it was solved with the aforementioned PRs.