Skip to content

v0.27.0

Choose a tag to compare

View all tags
@sachinprasadhs sachinprasadhs released this 02 Apr 23:02
· 55 commits to master since this release
v0.27.0
964781a
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

New Models

  • Gemma 4: Multimodal architecture supporting Audio, Text, and Vision, featuring both Dense and Mixture-of-Experts (MoE) variants.
  • T5Gemma 2: Integration of the T5Gemma2 model and presets to the Hub.
  • Embedding Gemma 3: Presets and conversion scripts to load weights from HuggingFace to KerasHub.
  • VideoPrism: A powerful video understanding model using a factorized encoder design for spatial and temporal processing, supporting both video-only and multimodal vision-language tasks.
  • MetaCLIP 2: Open-source implementation of CLIP by Meta.

New Features

  • TF-free Tokenizers: Pure Python/Keras implementations of SentencePieceTokenizer, BytePairTokenizer, StartEndPacker, and MultiSegmentPacker, removing TensorFlow dependencies for lighter deployments.

Export to Safetensors

  • Sequential Streaming Optimization: Optimized PyTorch memory overhead during Safetensors export via sequential streaming.
  • GPT-2 Safetensors: Migrated GPT-2 checkpoints from H5 to Hugging Face SafeTensors format.

Bug Fixes and Improvements

  • VGG Architecture: Fixed VGG architecture and updated checkpoints with activation fixes in FC layers.
  • OpenVino Backend: Fixed text generation (generate) for the OpenVino backend.

Contributors

We would like to thank our contributors for this release: @bermeitinger-b, @divyashreepathihalli, @hertschuh, @james77777778, @LakshmiKalaKadali, @laxmareddyp, @sachinprasadhs, @sineeli.

Full Changelog: v0.26.0...v0.27.0

Assets 2
Loading