Skip to content

Releases: huggingface/huggingface_hub

Patch release v0.11.1

28 Nov 10:48
@Wauplin Wauplin
v0.11.1
cc1b547
Compare
Choose a tag to compare
Patch release v0.11.1

Hot-fix to fix permission issues when downloading with hf_hub_download or snapshot_download. For more details, see #1220, #1141 and #1215.

Full changelog: v0.11.0...v0.11.1

Assets 2
Loading

Extended HfApi, pagination, simplified login and more

14 Nov 12:00
@Wauplin Wauplin
v0.11.0
fb9f637
Compare
Choose a tag to compare
Extended HfApi, pagination, simplified login and more

New features and improvements for HfApi

HfApi is the central point to interact with the Hub API (manage repos, create commits,...). The goal is to propose more and more git-related features using HTTP endpoints to allow users to interact with the Hub without cloning locally a repo.

Create/delete tags and branches

from huggingface_hub import create_branch, create_tag, delete_branch, delete_tag

create_tag(repo_id, tag="v0.11", tag_message="Release v0.11")
delete_tag(repo_id, tag="something") # If you created a tag by mistake

create_branch(repo_id, branch="experiment-154")
delete_branch(repo_id, branch="experiment-1") # Clean some old branches

Upload lots of files in a single commit

Making a very large commit was previously tedious. Files are now processed by chunks which makes it possible to upload 25k files in a single commit (and 1Gb payload limitation if uploading only non-LFS files). This should make it easier to upload large datasets.

  • Create commit by streaming a ndjson payload (allow lots of file in single commit) by @Wauplin in #1117

Delete an entire folder

from huggingface_hub import CommitOperationDelete, create_commit, delete_folder

# Delete a single folder
delete_folder(repo_id=repo_id, path_in_repo="logs/")

# Alternatively, use the low-level `create_commit`
create_commit(
    repo_id,
    operations=[
        CommitOperationDelete(path_in_repo="old_config.json") # Delete a file
        CommitOperationDelete(path_in_repo="logs/") # Delete a folder
    ],
    commit_message=...,
)

Support pagination when listing repos

In the future, listing models, datasets and spaces will be paginated on the Hub by default. To avoid breaking changes, huggingface_hub follows already pagination. Output type is currently a list (deprecated), will become a generator in v0.14.

  • Add support for pagination in list_models list_datasets and list_spaces by @Wauplin #1176
  • Deprecate output in list_models by @Wauplin in #1143

Misc

  • Allow create PR against non-main branch by @Wauplin in #1168
  • 1162 Reorder operations correctly in commit endpoint by @Wauplin in #1175

Login, tokens and authentication

Authentication has been revisited to make it as easy as possible for the users.

Unified login and logout methods

from huggingface_hub import login, logout

# `login` detects automatically if you are running in a notebook or a script
# Launch widgets or TUI accordingly
login()

# Now possible to login with a hardcoded token (non-blocking)
login(token="hf_***")

# If you want to bypass the auto-detection of `login`
notebook_login()  # still available
interpreter_login()  # to login from a script

# Logout programmatically
logout()
# Still possible to login from CLI
huggingface-cli login

Set token only for a HfApi session

from huggingface_hub import HfApi

# Token will be sent in every request but not stored on machine
api = HfApi(token="hf_***")

Stop using use_auth_token in favor of token, everywhere

token parameter can now be passed to every method in huggingface_hub. use_auth_token is still accepted where it previously existed but the mid-term goal (~6 months) is to deprecate and remove it.

  • Replace use_auth_token arg by token everywhere by @Wauplin in #1122

Respect git credential helper from the user

Previously, token was stored in the git credential store. Can now be in any helper configured by the user -keychain, cache,...-.

  • Refactor git credential handling in login workflow by @Wauplin in #1138

Better error handling

Helper to dump machine information

# Dump all relevant information. To be used when reporting an issue.
➜ huggingface-cli env

Copy-and-paste the text below in your GitHub issue.

- huggingface_hub version: 0.11.0.dev0
- Platform: Linux-5.15.0-52-generic-x86_64-with-glibc2.35
- Python version: 3.10.6
...

Misc

Modelcards

Few improvements/fixes in the modelcard module:

Cache assets

New feature to provide a path in the cache where any downstream library can store assets (processed data, files from the web, extracted data, rendered images,...)

  • [RFC] Proposal for a way to cache files in downstream libraries by @Wauplin in #1088

Documentation updates

Breaking changes

  • Cannot provide an organization to create_repo
  • identical_ok removed in upload_file
  • Breaking changes in arguments for validate_preupload_info, prepare_commit_payload, _upload_lfs_object (internal helpers for the commit API)
  • huggingface_hub.snapshot_download is not exposed as a public module anymore

Deprecations

  • Remove deprecated code from v0.9, v0.10 and v0.11 by @Wauplin in #1092
  • Rename languages to langage + remove duplicate code in tests by @Wauplin in #1169
  • Deprecate output in list_models by @Wauplin in #1143
  • Set back feature to create a repo when using clone_from by @Wauplin in #1187

Internal

Bugfixes & small improvements

Contributors

singingwolfboy, Narsil, and 9 other contributors
Assets 2
Loading

Patch release v0.10.1

11 Oct 07:48
@Wauplin Wauplin
v0.10.1
c87d402
Compare
Choose a tag to compare
Patch release v0.10.1

Hot-fix to force utf-8 encoding in modelcards. See #1102 and skops-dev/skops#162 (comment) for context.

Full Changelog: v0.10.0...v0.10.1

Assets 2
Loading

v0.10.0: Modelcards, cache management and more

28 Sep 07:26
@Wauplin Wauplin
v0.10.0
583d9f7
Compare
Choose a tag to compare
v0.10.0: Modelcards, cache management and more

Modelcards

Contribution from @nateraw to integrate the work done on Modelcards and DatasetCards (from nateraw/modelcards) directly in huggingface_hub.

>>> from huggingface_hub import ModelCard

>>> card = ModelCard.load('nateraw/vit-base-beans')
>>> card.data.to_dict()
{'language': 'en', 'license': 'apache-2.0', 'tags': ['generated_from_trainer', 'image-classification'],...}

Related commits

Related documentation

Cache management (huggingface-cli scan-cache and huggingface-cli delete-cache)

New commands in huggingface-cli to scan and delete parts of the cache. Goal is to manage the cache-system the same way for any dependent library that uses huggingface_hub. Only the new cache-system format is supported.

➜ huggingface-cli scan-cache
REPO ID                     REPO TYPE SIZE ON DISK NB FILES LAST_ACCESSED LAST_MODIFIED REFS                LOCAL PATH
--------------------------- --------- ------------ -------- ------------- ------------- ------------------- -------------------------------------------------------------------------
glue                        dataset         116.3K       15 4 days ago    4 days ago    2.4.0, main, 1.17.0 /home/wauplin/.cache/huggingface/hub/datasets--glue
google/fleurs               dataset          64.9M        6 1 week ago    1 week ago    refs/pr/1, main     /home/wauplin/.cache/
(...)

Done in 0.0s. Scanned 6 repo(s) for a total of 3.4G.
Got 1 warning(s) while scanning. Use -vvv to print details.

Related commits

Related documentation

Better error handling (and http-related stuff)

HTTP calls to the Hub have been harmonized to behave the same across the library.

Major differences are:

  • Unified way to handle HTTP errors using hf_raise_for_status (more informative error message)
  • Auth token is always sent by default when a user is logged in (see documentation).
  • package versions are sent as user-agent header for telemetry (python, huggingface_hub, tensorflow, torch,...). It was already the case for hf_hub_download.

Related commits

  • Always send the cached token when user is logged in by @Wauplin in #1064
  • Add user agent to all requests with huggingface_hub version (and other) by @Wauplin in #1075
  • [Repository] Add better error message by @patrickvonplaten in #993
  • Clearer HTTP error messages in huggingface_hub by @Wauplin in #1019
  • Handle backoff on HTTP 503 error when pushing repeatedly by @Wauplin in #1038

Breaking changes

  1. For consistency, the return type of create_commit has been modified. This is a breaking change, but we hope the return type of this method was never used (quite recent and niche output type).
  • Return more information in create_commit output by @Wauplin in #1066
  1. Since repo_id is now validated using @validate_hf_hub_args (see below), a breaking change can be caused if repo_id was previously miused. A HFValidationError is now raised if repo_id is not valid.

Miscellaneous improvements

Add support for autocomplete

http-based push_to_hub_fastai

  • Add changes for push_to_hub_fastai to use the new http-based approach. by @nandwalritik in #1040

Check if a file is cached

  • try_to_load_from_cache returns cached non-existence by @sgugger in #1039

Get file metadata (commit hash, etag, location) without downloading

  • Add get_hf_file_metadata to fetch metadata from the Hub by @Wauplin in #1058

Validate arguments using @validate_hf_hub_args

  • Add validator for repo id + decorator to validate arguments in huggingface_hub by @Wauplin in #1029
  • Remove repo_id validation in hf_hub_url and hf_hub_download by @Wauplin in #1031

⚠️ This is a breaking change if repo_id was previously misused ⚠️

Related documentation:

Documentation updates

Deprecations

  • ENH Deprecate clone_from behavior by @merveenoyan in #952
  • 🗑 Deprecate token in read-only methods of HfApi in favor of use_auth_token by @SBrandeis in #928
  • Remove legacy helper 'install_lfs_in_userspace' by @Wauplin in #1059
  • 1055 deprecate private and repo type in repository class by @Wauplin in #1057

Bugfixes & small improvements

  • Consider empty subfolder as None in hf_hub_url and hf_hub_download by @Wauplin in #1021
  • enable http request retry under proxy by @MrZhengXin in #1022
  • Add securityStatus to ModelInfo object with default value None. by @Wauplin in #1026
  • 👽️ Add size parameter for lfsFiles when committing on the hub by @coyotte508 in #1048
  • Use /models/ path for api call to update settings by @Wauplin in #1049
  • Globally set git credential.helper to store in google colab by @Wauplin in #1053
  • FIX notebook login by @Wauplin in #1073

Windows-specific bug fixes

  • Fix default cache on windows by @thomwolf in #1069
  • Degraded but fully working cache-system when symlinks are not supported by @Wauplin in #1067
  • Check symlinks support per directory instead of globally by @Wauplin in #1077

Contributors

coyotte508, thomwolf, and 10 other contributors
Assets 2
Loading

Patch release v0.9.1

25 Aug 15:41
@Wauplin Wauplin
v0.9.1
7aadc32
Compare
Choose a tag to compare
Patch release v0.9.1

Hot-fix error message on gated repositories (#1015).

Context: https://huggingface.co/CompVis/stable-diffusion-v1-4 has been widely shared in the last days but since it's a gated-repo, lots of users are getting confused by the Authentification error received. Error message is now more detailed.

Full Changelog: v0.9.0...v0.9.1

Assets 2
Loading

v0.9.0: Community API and new `push_to_hub` mixins

23 Aug 12:22
@Wauplin Wauplin
v0.9.0
13da627
Compare
Choose a tag to compare
v0.9.0: Community API and new `push_to_hub` mixins

Community API

Huge work to programmatically interact with the community tab, thanks to @SBrandeis !
It is now possible to:

  • Manage discussions (create_discussion, create_pull_request, merge_pull_request, change_discussion_status, rename_discussion)
  • Comment on them (comment_discussion, edit_discussion_comment)
  • List them (get_repo_discussions, get_discussion_details)

See full documentation for more details.

HTTP-based push_to_hub mixins

push_to_hub mixin and push_to_hub_keras have been refactored to leverage the http-endpoint. This means pushing to the hub will no longer require to first download the repo locally. Previous git-based version is planned to be supported until v0.12.

Miscellaneous API improvements

  • ✨ parent_commit argument for create_commit and related functions by @SBrandeis in #916
  • Add a helpful error message when commit_message is empty in create_commit by @sgugger in #962
  • ✨ create_commit: more user-friendly errors on HTTP 400 by @SBrandeis in #963
  • ✨ Add files_metadata option to repo_info by @SBrandeis in #951
  • Add list_spaces to HfApi by @cakiki in #889

Miscellaneous helpers (advanced)

Filter which files to upload in upload_folder

  • Allowlist and denylist when uploading a folder by @Wauplin in #994

Non-existence of files in a repo is now cached

  • Cache non-existence of files or completeness of repo by @sgugger in #986

Progress bars can be globally disabled via the HF_HUB_DISABLE_PROGRESS_BARS env variable or using disable_progress_bars/enable_progress_bars helpers.

  • Add helpers to disable progress bars globally + tests by @Wauplin in #987

Use try_to_load_from_cache to check if a file is locally cached

Documentation updates

Bugfixes & small improvements

Internal

Contributors

julien-c, juliensimon, and 14 other contributors
Assets 2
Loading

v0.8.1: lazy loading, git-aware cache file layout, new create_commit

15 Jun 15:53
@LysandreJik LysandreJik
v0.8.1
1da0161
Compare
Choose a tag to compare
v0.8.1: lazy loading, git-aware cache file layout, new create_commit

Git-aware cache file layout

v0.8.1 introduces a new way of caching files from the Hugging Face Hub, to two methods: snapshot_download and hf_hub_download.
The new approach is extensively documented in the Documenting files guide and we recommend checking it out to get a better understanding of how caching works.

New create_commit API

A new create_commit API allows users to upload and delete several files at once using HTTP-based methods. You can read more about it in this guide. The following convenience methods were also introduced:

  • upload_folder: Allows uploading a local directory to a repo.
  • delete_file allows deleting a single file from a repo.

upload_file now uses create_commit under the hood.

create_commit also allows creating pull requests with a create_pr=True flag.

None of the methods rely on Git locally.

Lazy loading

All modules will now be lazy-loaded. This should drastically reduce the time it takes to import huggingface_hub as it will no longer load all soft dependencies.

Improvements and bugfixes

Contributors

lsb, julien-c, and 4 other contributors
Assets 2
Loading

v0.7.0: Repocard metadata

30 May 12:18
@LysandreJik LysandreJik
v0.7.0
3491bf4
Compare
Choose a tag to compare
v0.7.0: Repocard metadata

Repocard metadata

This PR adds a metadata_update function that allows the user to update the metadata in a repository on the hub. The function accepts a dict with metadata (following the same pattern as the YAML in the README) and behaves as follows for all top level fields except model-index.

Examples:

Starting from

existing_results = [{
    'dataset': {'name': 'IMDb', 'type': 'imdb'},
    'metrics': [{'name': 'Accuracy', 'type': 'accuracy', 'value': 0.995}],
     'task': {'name': 'Text Classification', 'type': 'text-classification'}
}]

1. Overwrite existing metric value in existing result

new_results = deepcopy(existing_results)
new_results[0]["metrics"][0]["value"] = 0.999
_update_metadata_model_index(existing_results, new_results, overwrite=True)
[{'dataset': {'name': 'IMDb', 'type': 'imdb'},
  'metrics': [{'name': 'Accuracy', 'type': 'accuracy', 'value': 0.999}],
  'task': {'name': 'Text Classification', 'type': 'text-classification'}}]

2. Add new metric to existing result

new_results = deepcopy(existing_results)
new_results[0]["metrics"][0]["name"] = "Recall"
new_results[0]["metrics"][0]["type"] = "recall"
[{'dataset': {'name': 'IMDb', 'type': 'imdb'},
  'metrics': [{'name': 'Accuracy', 'type': 'accuracy', 'value': 0.995},
              {'name': 'Recall', 'type': 'recall', 'value': 0.995}],
  'task': {'name': 'Text Classification', 'type': 'text-classification'}}]

3. Add new result

new_results = deepcopy(existing_results)
new_results[0]["dataset"] = {'name': 'IMDb-2', 'type': 'imdb_2'}
[{'dataset': {'name': 'IMDb', 'type': 'imdb'},
  'metrics': [{'name': 'Accuracy', 'type': 'accuracy', 'value': 0.995}],
  'task': {'name': 'Text Classification', 'type': 'text-classification'}},
 {'dataset': ({'name': 'IMDb-2', 'type': 'imdb_2'},),
  'metrics': [{'name': 'Accuracy', 'type': 'accuracy', 'value': 0.995}],
  'task': {'name': 'Text Classification', 'type': 'text-classification'}}]

Improvements and bug fixes

Contributors

leondz, lvwerra, and merveenoyan
Assets 2
Loading

v0.6.0: fastai support, binary file support, skip LFS files when pushing to the hub

09 May 20:11
@LysandreJik LysandreJik
v0.6.0
a8b6f14
Compare
Choose a tag to compare
v0.6.0: fastai support, binary file support, skip LFS files when pushing to the hub

Disclaimer: This release was initially released with advertised support for #844. It was not released in this release and will be in v0.7.

fastai support

v0.6.0 introduces downstream (download) and upstream (upload) support for the fastai libraries. It supports fastai versions above 2.4.
The integration is detailed in the following blog.

  • Add fastai upstream and downstream capacities for fastai>=2.4 and fastcore>=1.3.27 versions by @omarespejel in #678

Automatic binary file tracking in Repository

Binary files are now rejected by default by the Hub. v0.6.0 introduces automatic binary file tracking through the auto_lfs_track argument of the Repository.git_add method. It also introduces the Repository.auto_track_binary_files method which can be used independently of other methods.

skip_lfs_file is now added to mixins

The parameter skip_lfs_files is now added to the different mixins. This will enable pushing files to the hub without first downloading the files above 10MB. This should drammatically reduce the time needed when updating a modelcard, a configuration file, and others.

  • ✨ add skip_lfs_files to mixins' push_to_hub by @nateraw in #858

Keras support improvement

The support for Keras model is greatly improved through several additions:

  • The save_pretrained_keras method now accepts a list of tags that will automatically be added to the repository.
  • Download statistics are now available on Keras models

Bugfixes and improvements

Contributors

leondz, julien-c, and 9 other contributors
Assets 2
Loading

v0.5.1: Patch release

07 Apr 19:10
@LysandreJik LysandreJik
v0.5.1
f81a0e9
Compare
Choose a tag to compare
v0.5.1: Patch release

This is a patch release fixing a breaking backward compatibility issue.

Linked PR: #822

Assets 2
Loading