2.1.0

🚀 What's New?

• Chronos-2 can now be deployed on AWS via SageMaker JumpStart. Please check this notebook for details.
• Scaled Dot Product Attention (SDPA) is now used as the default attention implementation in Chronos-2. If you need to use the previous eager implementation, please load the model with Chronos2Pipeline.from_pretrained(..., attn_implementation="eager").
• predict_df support has been added for older Chronos and Chronos-Bolt models. Now, all models (Chronos-2, Chronos-Bolt, Chronos) provide a unified pandas dataframe API. Note: Only Chronos-2 supports multivariate and covariate-informed forecasting.
• Chronos2Pipeline.embed has been added, enabling users to extract embeddings from the last layer of the Chronos-2 encoder.

🐛 Bug Fixes

• Fixed issues related to past-only covariates use during Chronos-2 fine-tuning. If you're fine-tuning Chronos-2 models, we strongly recommend upgrading to chronos-forecasting==2.1.0.
• Fixed issue related to multiple workers on windows.

All Changes

• Fix pandas install instruction by @abdulfatir in #326
• Scope down GitHub token permissions by @AdnaneKhan in #328
• Update Chronos-2 notebook install instructions by @abdulfatir in #333
• Add logging steps in fine-tuning example by @abdulfatir in #334
• [chronos-2] add support for SDPA by @kashif in #331
• Treat notebooks as documentation in linguist by @abdulfatir in #338
• Count notebooks as python in language stats by @abdulfatir in #347
• Move SageMaker notebook to a new path by @shchur in #348
• Revert "Move SageMaker notebook to a new path" by @shchur in #349
• fix validate_and_prepare_single_dict_task by separating "past only" and "future known" cov keys by @HarvestStars in #344
• Add Chronos-2 SageMaker notebook by @shchur in #350
• Update model link to Hugging Face by @shchur in #353
• Remove logo from README by @shchur in #355
• Update Notebooks and README by @abdulfatir in #356
• Add FutureWarning for CloudFront by @abdulfatir in #357
• Relax transformers lower bound to >=4.41 by @abdulfatir in #364
• Set num_workers=0 in Chronos-2 test data loader by @IliasAarab in #365
• Update long horizon heuristic for Chronos-Bolt by @abdulfatir in #366
• Add predict_df support for Chronos and Chronos-Bolt by @abdulfatir in #371
• Add Chronos2Pipeline.embed by @abdulfatir in #361
• Mask past-only covariates during loss computation by @abdulfatir in #379
• Chronos-2: Change default fine-tuning learning rate and remove experimental label by @abdulfatir in #381
• Update version to 2.1.0rc1 for pre-release by @abdulfatir in #383
• Update torch dependency version to >=2.2,<3 by @abdulfatir in #388
• Bump version to 2.1.0 by @abdulfatir in #389

New Contributors

• @AdnaneKhan made their first contribution in #328
• @kashif made their first contribution in #331
• @HarvestStars made their first contribution in #344
• @IliasAarab made their first contribution in #365

Full Changelog: v2.0.1...v2.1.0