Change8

v0.29.0

📦 huggingface-hubView on GitHub →
6 features🐛 9 fixes1 deprecations🔧 2 symbols

Summary

This release significantly expands serverless inference capabilities by adding official support for Fireworks AI, Hyperbolic, Nebius AI Studio, and Novita providers. It also includes several quality-of-life improvements and deprecates some legacy hf-inference specific features.

Migration Steps

  1. If you relied on deprecated hf-inference specific features like overriding model tasks (except for sentence-similarity), update your code to remove these calls.
  2. If you were using the `extra_parameters` argument in `InferenceClient`, rename it to `extra_body`.

✨ New Features

  • Added official support for four new serverless Inference Providers on the Hugging Face Hub: Fireworks AI, Hyperbolic, Nebius AI Studio, and Novita.
  • Added support for YuE (music generation) from fal.ai.
  • Defaulted to `base_url` if provided in inference settings.
  • Added dynamic inference providers mapping to InferenceClient.
  • Added `HF_DEBUG` environment variable for debugging/reproducibility.
  • Added new environment variable `HF_HUB_USER_AGENT_ORIGIN` to set the origin of calls in the user-agent.

🐛 Bug Fixes

  • Fixed automatic-speech-recognition output parsing.
  • Replaced `urljoin` for `HF_ENDPOINT` paths.
  • Ensured staging token is not sent to production.
  • Fixed curlify utility.
  • Improved `whoami()` error messages by specifying the token source.
  • Fixed error message when an invalid token is used during file download.
  • Fixed `test_dataset_info` due to a missing dummy dataset.
  • Fixed `is_jsonable` when an integer key is present in a dictionary.
  • Fixed Inference Client VCR tests.

Affected Symbols

⚡ Deprecations

  • Deprecation of some hf-inference specific features: wait-for-model header, inability to override model's task, get_model_status, and list_deployed_models. (Note: task override for sentence-similarity was partially reverted/allowed).