2.3.0

What's Changed

🚀 Highlights

• New cloud deployment guide: run Chronos-2 on AWS in 3 lines of code
• Fine-tuning with Chronos2Pipeline.fit() now supports larger-than-memory datasets
• New efficient preprocessing module chronos.chronos2.preprocess resulting in up to 20x faster input preprocessing for Chronos-2
• Support for transformers>=5 and pandas>=3

❗Deprecations & breaking changes

• When providing inputs to Chronos2Pipeline as a list[dict], all dictionaries must share the same schema (same number of target dimensions and same covariate keys). To forecast inputs with different schemas, loop over them and call the model once per schema.

  # ❌ Different covariate keys across items — no longer allowed
  pipeline.predict([
      {"target": [...], "past_covariates": {"price": [...], "promo": [...]}},
      {"target": [...], "past_covariates": {"price": [...]}},
  ])
  
  # ❌ Mixed univariate and multivariate targets — no longer allowed
  pipeline.predict([
      {"target": np.zeros(100)},       # shape (history_length,)
      {"target": np.zeros((3, 100))},  # shape (n_variates, history_length)
  ])

• Following methods have been deprecated. Please use chronos.chronos2.preprocess.from_* and helpers from chronos.df_utils instead.
  • Input preparation methods in chronos.chronos2.dataset:.
    • Chronos2Dataset.convert_inputs
    • validate_and_prepare_single_dict_task
    • convert_list_of_tensors_input_to_list_of_dicts_input
    • convert_tensor_input_to_list_of_dicts_input
    • convert_fev_window_to_list_of_dicts_input
  • DataFrame utilities in chronos.df_utils:
    • convert_df_input_to_list_of_dicts_input
    • validate_df_inputs
• Removed dependency on scikit-learn

✨ New features

• Add support for transformers>=5.0 by @kashif in #416
• New module chronos.chronos2.preprocess with efficient preprocessing methods for Chronos-2 by @shchur in #466, #493, #506
• Public API of Chronos2Pipeline now accepts preprocessed data (e.g. backed by a memory-mapped Hugging Face datasets.Dataset) by @shchur in #507, #512

🐛 Bugfixes & improvements

• Allow explicitly passing the frequency to pipeline.predict_df by @shchur in #449
• Fix XDG_CACHE_HOME typo in boto_utils by @shaun0927 in #488
• Defer boto3 import and raise actionable ImportError for s3:// loads by @shaun0927 in #489
• Fix pip install command for chronos-forecasting by @Arseni1919 in #481
• Minor fixes and improvements by @abdulfatir, @shchur in #467, #505, #510, #511,

📚 Documentation

• Describe deployment options via AutoGluon-Cloud by @shchur in #503
• Update notebook on Chronos-2 deployment to SageMaker by @shchur in #444

New Contributors

• @Arseni1919 made their first contribution in #481
• @shaun0927 made their first contribution in #488

Full Changelog: v2.2.2...v2.3.0