TensorFlow 2.18 Release: Key Features and Updates You Need to Know
TensorFlow, the popular open-source machine learning framework developed by Google, has officially announced the release of TensorFlow 2.18. This update brings significant improvements and new features that enhance usability, performance, and compatibility with the latest tools in the data science ecosystem. In this article, we’ll explore the highlights of this release, including support for NumPy 2.0, updates to LiteRT, improvements in CUDA support, and more.
Key Updates in TensorFlow 2.18
Support for NumPy 2.0
One of the standout features of TensorFlow 2.18 is its compatibility with NumPy 2.0. This update marks a significant milestone in TensorFlow’s evolution, as it allows developers to leverage the new capabilities of NumPy while ensuring that the majority of TensorFlow APIs work seamlessly with it. However, users should be aware that there may be some edge cases that could lead to issues, such as out-of-boundary conversion errors and numpy scalar representation errors.
The changes to NumPy’s type promotion rules (as outlined in NEP 50) could affect the precision of computations, potentially resulting in type errors or numerical discrepancies in results. To assist developers in navigating these changes, TensorFlow has updated several tensor APIs for compatibility with NumPy 2.0, while retaining the out-of-boundary conversion behavior familiar to users of NumPy 1.x.
For those migrating to this new version, it’s advisable to refer to the comprehensive NumPy 2 migration guide to ensure a smooth transition.
Transition to LiteRT
In a move to streamline its development process, TensorFlow is transitioning TFLite to a new codebase called LiteRT. This change is expected to enhance the functionality and performance of TensorFlow Lite. The development team plans to gradually migrate TFLite’s existing codebase to LiteRT over the coming months.
More Read
Once the migration is complete, contributions will be accepted directly through the LiteRT repository, and binary TFLite releases will be phased out. Developers are encouraged to switch to LiteRT for the latest updates and features in lightweight model deployment. This transition underscores TensorFlow’s commitment to evolving its tools and adapting to the needs of its users.
Hermetic CUDA for Reproducible Builds
TensorFlow 2.18 introduces Hermetic CUDA, which enhances the reliability of building TensorFlow from source. With this update, Bazel will automatically download specific versions of CUDA, CUDNN, and NCCL distributions, using these as dependencies in various Bazel targets. This means that developers can achieve more reproducible builds, which is particularly beneficial for Google ML projects that require consistent performance and compatibility across different environments.
By reducing the dependency on locally installed CUDA versions, TensorFlow aims to simplify the development workflow and minimize potential errors that could arise from version mismatches. For more detailed information on implementing Hermetic CUDA, developers can refer to the official documentation.
Enhanced CUDA Support
The CUDA update in TensorFlow 2.18 is particularly exciting for users with NVIDIA GPUs. TensorFlow now ships with dedicated CUDA kernels optimized for GPUs with a compute capability of 8.9, improving performance on popular Ada-Generation GPUs, including the NVIDIA RTX 40** series, L4, and L40.
Additionally, TensorFlow has made the decision to stop shipping CUDA kernels for older compute capabilities (specifically, 5.0). As a result, the oldest NVIDIA GPU generation supported by precompiled Python packages is now the Pascal generation, which has a compute capability of 6.0. Users with Maxwell GPUs are advised to stick with TensorFlow version 2.16 or consider compiling TensorFlow from source, provided the CUDA version in use still supports these older GPUs.
Transitioning to Keras 3.0
Another noteworthy update in the TensorFlow ecosystem is the ongoing development of Keras, which is transitioning to a new multi-backend architecture with Keras 3.0. This change is expected to enhance the flexibility and capability of Keras, making it even more powerful for building deep learning models. Updates on these developments will be published on keras.io, providing users with the latest information and resources as they adapt to the new framework.
Conclusion
TensorFlow 2.18 brings a host of new features and updates that promise to enhance user experience and boost performance across various applications. With support for NumPy 2.0, the transition to LiteRT, Hermetic CUDA for reproducible builds, improved CUDA support, and exciting developments in Keras, TensorFlow continues to solidify its position as a leading tool in the machine learning landscape. As always, developers are encouraged to explore the full release notes and migration guides to leverage these updates effectively.
Source: Original Article
