The v1.0 release is a major milestone for the huggingface_hub library. It marks our commitment to API stability and the maturity of the library. We have made several improvements and breaking changes to make the library more robust and easier to use.
This guide is intended to help you migrate your existing code to the new version. If you have any questions or feedback, please let us know by opening an issue on GitHub.
huggingface_hub now requires Python 3.9 or higher. Python 3.8 is no longer supported.
The huggingface_hub library now uses httpx instead of requests for HTTP requests. This change was made to improve performance and to support both synchronous and asynchronous requests the same way. We therefore dropped both requests and aiohttp dependencies.
This is a major change that affects the entire library. While we have tried to make this change as transparent as possible, you may need to update your code in some cases. Here is a list of breaking changes introduced in the process:
HTTP_PROXY and HTTPS_PROXY environment variables.configure_http_backend function has been removed. You should now use set_client_factory() and set_async_client_factory() to configure the HTTP clients.requests.HTTPError anymore, but from httpx.HTTPError. We recommend catching huggingface_hub.HfHubHttpError which is a subclass of requests.HTTPError in v0.x and of httpx.HTTPError in v1.x. Catching from the huggingface_hub error ensures your code is compatible with both the old and new versions of the library.httpx does not have the concept of SSLError. It is now a generic httpx.ConnectError.LocalEntryNotFoundError: This error no longer inherits from HTTPError. We now define a EntryNotFoundError (new) that is inherited by both LocalEntryNotFoundError (if file not found in local cache) and RemoteEntryNotFoundError (if file not found in repo on the Hub). Only the remote error inherits from HTTPError.InferenceClient: The InferenceClient can now be used as a context manager. This is especially useful when streaming tokens from a language model to ensure that the connection is closed properly.AsyncInferenceClient: The trust_env parameter has been removed from the AsyncInferenceClient’s constructor. Environment variables are trusted by default by httpx. If you explicitly don’t want to trust the environment, you must configure it with set_client_factory().For more details, you can check PR #3328 that introduced httpx.
The migration from requests to httpx brings several key improvements that enhance the library’s performance, reliability, and maintainability:
Thread Safety and Connection Reuse: httpx is thread-safe by design, allowing us to safely reuse the same client across multiple threads. This connection reuse reduces the overhead of establishing new connections for each HTTP request, improving performance especially when making frequent requests to the Hub.
HTTP/2 Support: httpx provides native HTTP/2 support, which offers better efficiency when making multiple requests to the same server (exactly our use case). This translates to lower latency and reduced resource consumption compared to HTTP/1.1.
Unified Sync/Async API: Unlike our previous setup with separate requests (sync) and aiohttp (async) dependencies, httpx provides both synchronous and asynchronous clients with identical behavior. This ensures that InferenceClient and AsyncInferenceClient have consistent functionality and eliminates subtle behavioral differences that previously existed between the two implementations.
Improved SSL Error Handling: httpx handles SSL errors more gracefully, making debugging connection issues easier and more reliable.
Future-Proof Architecture: httpx is actively maintained and designed for modern Python applications. In contrast, requests is in maintenance mode and won’t receive major updates like thread-safety improvements or HTTP/2 support.
Better Environment Variable Handling: httpx provides more consistent handling of environment variables across both sync and async contexts, eliminating previous inconsistencies where requests would read local environment variables by default while aiohttp would not.
The transition to httpx positions huggingface_hub with a modern, efficient, and maintainable HTTP backend. While most users should experience seamless operation, the underlying improvements provide better performance and reliability for all Hub interactions.
Now that all repositories on the Hub are Xet-enabled and that hf_xet is the default way to download/upload files, we’ve removed support for the hf_transfer optional package. The HF_HUB_ENABLE_HF_TRANSFER environment variable is therefore ignored. Use HF_XET_HIGH_PERFORMANCE instead.
The Repository class has been removed in v1.0. It was a thin wrapper around the git CLI for managing repositories. You can still use git directly in the terminal, but the recommended approach is to use the HTTP-based API in the huggingface_hub library for a smoother experience, especially when dealing with large files.
Here is a mapping from the legacy Repository class to the new HfApi one:
Repository method | HfApi method |
|---|---|
repo.clone_from | snapshot_download |
repo.git_add + git_commit + git_push | upload_file(), upload_folder(), create_commit() |
repo.git_tag | create_tag |
repo.git_branch | create_branch |
HfFolder was used to manage the user access token. Use login() to save a new token, logout() to delete it and whoami() to check the user associated to the current token. Finally, use get_token() to retrieve user’s token in a script.
InferenceApi was a class to interact with the Inference API. It is now recommended to use the InferenceClient class instead.
Some methods and parameters have been removed in v1.0. The ones listed below have already been deprecated with a warning message in v0.x.
constants.hf_cache_home has been removed. Please use HF_HOME instead.use_auth_token parameters have been removed from all methods. Please use token instead.get_token_permission method has been removed.update_repo_visibility method has been removed. Please use update_repo_settings instead.is_write_action parameter has been removed from build_hf_headers as well as write_permission from login. The concept of “write permission” has been removed and is no longer relevant now that fine-grained tokens are the recommended approach.new_session parameter in login has been renamed to skip_if_logged_in for better clarity.resume_download, force_filename, and local_dir_use_symlinks parameters have been removed from hf_hub_download and snapshot_download.library, language, tags, and task parameters have been removed from list_models.Cache management from the CLI has been redesigned to follow a Docker-inspired workflow. The deprecated huggingface-cli has been removed, hf (introduced in v0.34) replaces it with a clearer ressource-action CLI.
The legacy hf cache scan and hf cache delete commands are also removed in v1.0 and are replaced with the new trio below:
hf cache ls lists cache entries with concise table, JSON, or CSV output. Use --revisions to inspect individual revisions, add --filter expressions such as size>1GB or accessed>30d, and combine them with --quiet when you only need the identifiers.hf cache rm deletes selected cache entries. Pass one or more repo IDs (for example model/bert-base-uncased) or revision hashes, and optionally add --dry-run to preview or --yes to skip the confirmation prompt. This replaces both the interactive TUI and --disable-tui workflows from the previous command.hf cache prune performs the common cleanup task of deleting unreferenced revisions in one shot. Add --dry-run or --yes in the same way as with hf cache rm.Finally, the [cli] extra has been removed - The CLI now ships with the core huggingface_hub package.
All TensorFlow-related code and dependencies have been removed in v1.0. This includes the following breaking changes:
huggingface_hub[tensorflow] is no longer a supported extra dependencysplit_tf_state_dict_into_shards and get_tf_storage_size utility functions have been removed.tensorflow, fastai, and fastcore versions are no longer included in the built-in headers.The Keras 2.x integration has also been removed. This includes the KerasModelHubMixin class and the save_pretrained_keras, from_pretrained_keras, and push_to_hub_keras utilities. Keras 2.x is a legacy and unmaintained library. The recommended approach is to use Keras 3.x which is tightly integrated with the Hub (i.e. it contains built-in method to load/push to Hub). If you still want to work with Keras 2.x, you should downgrade huggingface_hub to v0.x version.
The upload_file() and upload_folder() functions now return the URL of the commit created on the Hub. Previously, they returned the URL of the file or folder. This is to align with the return value of create_commit(), delete_file() and delete_folder().
Update on GitHub