Pipeline has no attribute '_execution_device'

[choidaedae]

Describe the bug

Hello, I implemented my own custom pipeline referring StableDiffusionPipeline (RepDiffusionPipeline), but there are some issues
I called "accelerator.prepare" properly, and mapped the models on device (with "to.(accelerator.device)")
But when I call pipeline and the 'call' function is called, sometimes I met the error
It is not only problem in using multi-gpu, it occurs when I use single gpu.
For example, I defined my pipeline for my validation in training code like this:

val_pipe = RepDiffusionPipeline.from_pretrained(
                        "runwayml/stable-diffusion-v1-5",
                        unet=accelerator.unwrap_model(unet),
                        rep_encoder=accelerator.unwrap_model(rep_encoder),
                        vae=accelerator.unwrap_model(vae),
                        revision=None, variant=None, torch_dtype=weight_dtype, safety_checker=None
                    ).to(accelerator.device)

then, when I called 'val_pipe' like this:

model_pred = val_pipe(
                                        image = condition_original_image if args.val_mask_op else data["original_images"],
                                        representation = representation,
                                        prompt = "",
                                        num_inference_steps = 20,
                                        image_guidance_scale = 1.5,
                                        guidance_scale = scale,
                                        generator = generator
                                    ).images[0]

At that time, the error "RepDiffusionPipeline has no attribute '_execution_device'" occurs. (Not always, just randomly)
How can I solve this issue, or what part of my code can be doubted and fixed?
Thank you for reading:)

Reproduction

It occurs randomly, so there is no option to reproduce...

But when I call the defined pipeline, it occurs randomly.

Logs

RepDiffusionPipeline has no attribute '_execution_device'

System Info

I tried to test in various diffusers & python versions, but the problem still occurs.
In now, I am running my code in diffusers 0.27.2, python 3.10.14.

WARNING[XFORMERS]: xFormers can't load C++/CUDA extensions. xFormers was built for:
PyTorch 2.2.2+cu121 with CUDA 1201 (you have 2.2.2+cu118)
Python 3.10.14 (you have 3.10.14)
Please reinstall xformers (see https://github.com/facebookresearch/xformers#installing-xformers)
Memory-efficient attention, SwiGLU, sparse and more won't be available.
Set XFORMERS_MORE_DETAILS=1 for more details

Copy-and-paste the text below in your GitHub issue and FILL OUT the two last points.

• diffusers version: 0.27.2
• Platform: Linux-5.4.0-132-generic-x86_64-with-glibc2.31
• Python version: 3.10.14
• PyTorch version (GPU?): 2.2.2+cu118 (True)
• Huggingface_hub version: 0.24.3
• Transformers version: 4.43.3
• Accelerate version: 0.33.0
• xFormers version: 0.0.25.post1
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help?

@sayakpaul @yiyixuxu

[yiyixuxu]

hi @choidaedae

would you be able to provide a minimum reproducible code example? i.e. a script I can run on my end to get the same error that you're getting?

it would really help us understand the problem :)

[choidaedae]

hi @yiyixuxu
Thanks for replying.
It occurs when I validate with my customized pipeline, in training code of my private repos.
Can I invite you to my private repository and can you run the code?
If it can, I'll clean the code and data to run it easily and invite you.
Thanks :)

[LinB203]

same question.

[MuQY1818]

Same question

[yiyixuxu]

@LinB203 @MuQY1818
a minimum reproducible code example to help me understand the problem will be appreciated!

[luyvlei]

I had a similar problem, and it happened randomly so it was hard to reproduce, and I have now abandoned this attribute

[asomoza]

so can we assume this issue is resolved? OP also didn't provide any code to reproduce the error.

[choidaedae]

I am sorry for not providing any code, but it occurs randomly and actually I'm not working with this pipeline now, so I can't provide it.

If anyone still has this problem, providing the code can be userful for huggingface.
I will let you know if a similar error occurs again later.

Thank you for your hard work : )

[yiyixuxu]

think we can try to catch the exception for that property - so maybe in the future it happens we will have some visibility into the problem

[Lum1104]

Same problem. This bug will disappear if I run it again.

[myendless1]

Same problem, occurs randomly.

[Arlaz]

Same problem, occurs about half of the time with the same script