Stable Diffusion 2-1 Not Loaded on Colab with Minimal Loading

[maliozer]

Describe the bug

Within this simple loading script:

from diffusers import DiffusionPipeline

pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1")

The progress is stucked there is no error at all,
Fetching 13 files: 0% 0/13 [05:54<?, ?it/s]

Reproduction

Colab: https://colab.research.google.com/drive/1tzys4nVo95JAJgd_0MWp4s9Nm71wxQlZ?usp=sharing

Logs

Fetching 13 files: 0%
0/13 [05:54<?, ?it/s]
tokenizer/special_tokens_map.json: 100%
460/460 [00:00<00:00, 9.94kB/s]
tokenizer/tokenizer_config.json: 100%
824/824 [00:00<00:00, 50.3kB/s]

System Info

Colab

• diffusers version: 0.25.0
• Platform: Linux-6.1.58+-x86_64-with-glibc2.35
• Python version: 3.10.12
• PyTorch version (GPU?): 2.1.0+cu121 (False)
• Huggingface_hub version: 0.20.1
• Transformers version: 4.35.2
• Accelerate version: not installed
• xFormers version: not installed
• Using GPU in script?:
• Using distributed or parallel set-up in script?:

Who can help?

@yiyixuxu @DN6 @sayakpaul @patrickvonplaten

[sayakpaul]

Hmm this seems to be a problem only on Colab. I tried it on our internal server and on my local machine. It wasn't a problem.

[DN6]

The issue has also been raised on the Colab repo
googlecolab/colabtools#4272

[pranauv1]

It's very weird, but running it on a .py file seems to be working.

Download the model through a Python file

from diffusers import DiffusionPipeline
import torch

DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")

Save the above as a .py file and run it on Colab, the model will be cached!

Now call the pipeline
This won't attempt to download from HF since the model is already cached locally!

pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")

[ritwikraha]

I don't know if this is practical for everyone but running it on kaggle-notebooks seems to be working

https://www.kaggle.com/ritzraha/diffusion-template

So it is highly probable that this is a colab issue, FYI I am still having the same problem (as documented above) on colab

[imanphung]

pipe = StableDiffusionControlNetPipeline.from_pretrained( "SG161222/Realistic_Vision_V5.1_noVAE", controlnet=controlnet, torch_dtype=torch.float16, scheduler=noise_scheduler, vae=vae, token="", ).to("cuda")

I added the huggingface token inside then it works fine!
You can try it

[maliozer]

It may await to read hf_token from colab but never access

[KevinJBoyer]

@imanphung thank you! Confirming that adding token="" to the .from_pretrained call worked for me too.

[maliozer]

@imanphung thank you! Confirming that adding token="" to the .from_pretrained call worked for me too.

This is not worked for me, still need a valid solution

[math-rb]

It's very weird, but running it on a .py file seems to be working.

Download the model through a Python file

from diffusers import DiffusionPipeline
import torch

DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")

Save the above as a .py file and run it on Colab, the model will be cached!

Now call the pipeline This won't attempt to download from HF since the model is already cached locally!

pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")

Worked here.

[sayakpaul]

We're on top of it: huggingface/huggingface_hub#1953. Cc: @Wauplin

[Wauplin]

Thanks everyone for reporting and sorry for the inconvenience. I have been able to reproduce the issue and created a PR (huggingface/huggingface_hub#1953) to fix it. The plan is to make a hot-fix release once it's merged and make it the default in google colab. In the meantime, a temporary solution is to install a different version of huggingface_hub before executing any code. More details in huggingface/huggingface_hub#1953 (comment).

[Wauplin]

Update: huggingface_hub 0.20.2 is now the default version on Google Colab, meaning this issue should be fixed for everyone without having to upgrade (see googlecolab/colabtools#4272 (comment)).