Breaking changes
There are no breaking changes in this release. TF-Keras 2.22.0rc0 requires tensorflow>=2.22.0rc0,<2.23, and the minimum Python requirement remains >=3.10.
What's Changed
- Security Enhancement: Improved security of the cache directory fallback. d27a3bad9f
- Security Enhancement: Reject HDF5 datasets that read from the host or blow up memory when loading. #921
- Security Enhancement: Reject hard-link and symlink tar members whose destination or link target escapes the extraction directory. 844bcd4c55
- Functional Model Fix: Fixed deserialization of constant NumPy arrays and TF tensors that are arguments in a Functional model. 639424c379
- Raise
NotImplementedErrorforBatchNormalizationwithrenormunder a DTensor strategy. 9ab9f56224 - Test against TensorFlow 2.22 and align CUDA dependencies (cuDNN 9.10, NCCL 2.30, NVSHMEM) in CI. #936
- Test fixes for TF32 on H100 GPUs, Pillow >= 13, and stricter
name_scopeargument validation. #934 - Refresh API compatibility goldens. f8e51f8c17, afcb2544fb
Full Changelog: v2.21.0...v2.22.0rc0