v0.29.1

New Models

• Swin Transformer: Added Swin Transformer, a hierarchical vision transformer that uses shifted windows for efficient self-attention computation, enabling strong performance across image classification, object detection, and segmentation tasks.

• Gemma4: Added Gemma4, Google's latest lightweight open model with assistant-style capabilities and speculative decoding support for significantly faster text generation.

New Features

• HuggingFace Sync: Updated the model preset catalog with the latest weights and configurations to HuggingFace, expanding the range of ready-to-use pretrained models.

• ViT Presets: Refined Vision Transformer preset configurations with corrected naming conventions and accurate checkpoint mappings for improved usability.

Bug Fixes and Improvements

• Preprocessing:
  • Fixed convert_preprocessing_outputs to safely handle Python-native outputs and prevent .dtype errors.
• Tokenizers:
  • Fixed compatibility with tokenizers>=0.23.
• KaggleHub:
  • Excluded kagglehub 1.0.1 due to broken kagglesdk 0.1.24 compatibility.
• Attention Layers:
  • Fixed shape passed to inner EinsumDense in XLNET attention.
  • Fixed dynamic shape handling in Moonshine multi-head attention layer.
• Generation:
  • Fixed parallelization of input data in generate().
• Model Presets:
  • Fixed vit_encoder_layer name typo.
  • Fixed missing parameter counts in T5 preset metadata.
• Dependencies:
  • Updated Keras version requirement to 3.14.

Contributors

We would like to thank our contributors for this release: @laxmareddyp, @AdonaiVera, @dvadym, @hertschuh, @MaitrySinha21, @sachinprasadhs.

New Contributors

• @AdonaiVera made their first contribution
• @dvadym made their first contribution

Full Changelog: v0.29.0...v0.29.1