v0.28.0
Breaking Changes📦 huggingface-hubView on GitHub →
⚠ 1 breaking✨ 5 features🐛 10 fixes⚡ 1 deprecations🔧 9 symbols
Summary
This release introduces major updates to the InferenceClient, enabling unified inference across multiple third-party providers using Hugging Face Hub model IDs. Additionally, HfApi has been updated with new repository properties and the deprecated `like` endpoint has been removed.
⚠️ Breaking Changes
- The `like` endpoint in HfApi has been removed to prevent misuse. To remove existing likes, use the `unlike` endpoint instead.
Migration Steps
- If you were using the removed `like` endpoint, switch to using the `unlike` endpoint to remove existing likes.
✨ New Features
- The `InferenceClient` now supports third-party providers (like Replicate, Together AI, fal-ai) offering a unified interface for inference.
- Unified model IDs are enforced, always referencing Hugging Face Hub model IDs when using external providers.
- Added support for using Hugging Face Token for routing calls to third-party providers, enabling billing through the HF account.
- Added `text-to-video` task support to `InferenceClient`.
- HfApi now includes new repository properties: `usedStorage` and `resourceGroup`.
🐛 Bug Fixes
- Fixed an issue where `super_squash_history` revision was not URLencoded.
- Replaced usage of 'model repo' with 'repo' in docstrings for consistency.
- Fixed a 404 NOT FOUND issue caused by an endpoint tail slash.
- Fixed an issue with calling `typing.get_type_hints` on a `ModelHubMixin`.
- Fixed a typo in the codebase.
- Updated docstring for rejection reason.
- Added timeout to `WeakFileLock`.
- Fixed `CardData.get()` to correctly respect default values when the result is `None`.
- Fixed `RepoCard.load` when a repo_id that is also a directory path is passed.
- Fixed an issue causing filename too long errors when downloading to a local folder.
Affected Symbols
⚡ Deprecations
- The `logit_bias` parameter in `chat_completion()` within `InferenceClient` is now flagged as UNUSED.