Prepare for release 0.26

[hanouticelina]

PR following the 0.25.0 release to prepare for next one.

Main Changes

• bump version to 0.26.0.dev0
• remove url_to_filename, filename_to_url and cached_download functions.
• remove legacy_cache_layout argument from hf_hub_download.
• remove cached_download tests and legacy_cache_layout test.
• remove deprecate_positional_args from InferenceClient.__init__ and AsyncInferenceClient.__init__, This means enforcing keyword-only arguments for all parameters after model is now effective!

[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.

[Wauplin]

@LysandreJik's favourite type of PR with 600+ lines less to maintain! 😄 Removing cached_download is a big milestone. It shouldn't cause too many breaks in the wild given we've already took care of opening some PRs + deprecating it on a long cycle. Let's keep an eye open once it's released though :)

Thanks @hanouticelina for the cleaning. Let's wait for @LysandreJik review before merging

[LysandreJik]

Very happy to see this cleaned up!

[hanouticelina]

thanks for the reviews! let's merge it 🧹

[njbrake]

Hi! I'm here because it looks like removing the cache_download function in the newest release broke the AWS Sagemaker Huggingface Inference Toolkit.

I'm using the huggingface-pytorch-inference:2.1.0-transformers4.37.0-cpu-py310-ubuntu22.04 container which as far as I can tell is the newest container available.

https://github.com/aws/deep-learning-containers/blob/master/available_images.md#huggingface-inference-containers

Traceback (most recent call last):
  File "/usr/local/bin/dockerd-entrypoint.py", line 21, in <module>
    from sagemaker_huggingface_inference_toolkit import serving
  File "/opt/conda/lib/python3.10/site-packages/sagemaker_huggingface_inference_toolkit/serving.py", line 18, in <module>
    from sagemaker_huggingface_inference_toolkit import handler_service, mms_model_server
  File "/opt/conda/lib/python3.10/site-packages/sagemaker_huggingface_inference_toolkit/handler_service.py", line 28, in <module>
    from sagemaker_huggingface_inference_toolkit.transformers_utils import (
  File "/opt/conda/lib/python3.10/site-packages/sagemaker_huggingface_inference_toolkit/transformers_utils.py", line 26, in <module>
    from sagemaker_huggingface_inference_toolkit.diffusers_utils import get_diffusers_pipeline, is_diffusers_available
  File "/opt/conda/lib/python3.10/site-packages/sagemaker_huggingface_inference_toolkit/diffusers_utils.py", line 32, in <module>
    from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler, StableDiffusionPipeline
  File "/opt/conda/lib/python3.10/site-packages/diffusers/__init__.py", line 5, in <module>
    from .utils import (
  File "/opt/conda/lib/python3.10/site-packages/diffusers/utils/__init__.py", line 38, in <module>
    from .dynamic_modules_utils import get_class_from_dynamic_module
  File "/opt/conda/lib/python3.10/site-packages/diffusers/utils/dynamic_modules_utils.py", line 28, in <module>
    from huggingface_hub import HfFolder, cached_download, hf_hub_download, model_info
ImportError: cannot import name 'cached_download' from 'huggingface_hub' (/opt/conda/lib/python3.10/site-packages/huggingface_hub/__init__.py)

[Wauplin]

Hi @njbrake thanks for reporting and sorry for the inconvenience. I escalated this internally to see what can be done to fix it.
The error comes from the fact that diffusers tries to import cached_download which has been removed in huggingface_hub. However, we've already updated diffusers before to avoid this issue (see huggingface/diffusers#8419, released in diffusers 0.29.0). So updating diffusers to a recent version should fix it. Could you tell us which version of diffusers is installed in the container you are using?

[philschmid]

@njbrake could you share what you are doing? Do you have a requirements.txt which installs some dependencies?

[njbrake]

Hi, thanks all for the quick response. In my docker container, I have the following requirements.txt which updates the dependencies I care about. I am working in NLP and not using the diffusers library for anything, so I don't do anything to specify that it needs to be upgraded.

torch==2.0
transformers==4.41.2
nltk==3.8.1
spacy==3.4.4
peft==0.12.0

In order to get around the current situation it sounds like I can either pin the huggingface_hub version or enforce that the diffusers library gets upgraded? It seems weird to need to specify a version of the diffusers library when I don't use the diffusers library for anything in my application, it's only getting pulled in as a dependency pulled in by the SM HF inference toolkit.

Thanks!

[philschmid]

We are looking into this but it seems an update of transformers leads to a change of huggingface_hub, which removed a method which is used in a other library. As quick workaround can you add huggingface_hub<0.26.0 to your requirements.txt.