TensorFlow 2.18 switches to NumPy 2 and Hermetic CUDA
This is new: TensorFlow 2.18 integrates the current version 2.0 of NumPy and, with Hermetic CUDA, will no longer require local CUDA libraries during the build.
(Image: iX)
The recently released version 2.18 of TensorFlow now supports NumPy 2.0 and is compiled with it by default. It also uses Hermetic CUDA in the build and works on newer GPUs with a compute capability of 8.9 on the Nvidia grid.
According to the TensorFlow team, most APIs work smoothly when switching to NumPy 2, but in a few cases there are error messages, such as out-of-boundary conversion or numpy scalar representation. The team has deliberately retained the NumPy 1 behavior for Python conversion. Furthermore, developers should note the new type promotion rules of NumPy 2, which also have an impact on TensorFlow results and can lead to errors or less precise scalar results. The team only intends to continue support for NumPy 1.26 until 2025.
Videos by heise
CUDA no longer local
Developers who build TensorFlow from source will receive Hermetic CUDA in the future. TensorFlow will then load CUDNN and NCCL from the network so that CUDA is no longer installed locally. The TensorFlow team cites better reproducible builds for Google's own ML projects as the reason for this. The team has removed TensorRT from CUDA builds due to "code health improvement".
TensorFlow binaries now offer kernels for GPUs with a compute capability of 8.9 for Ada GPUs, which can be found on newer hardware such as NVIDIA RTX 40**, L4 or L40. However, GPUs prior to the Pascal generation (6.0) are no longer supported. The team recommends that users of older GPUs stick with version 2.16 or build TensorFlow themselves for their platform.
When the team has completed the conversion from TensorFlow Lite to RTLite by the end of the year, contributors will be able to participate, for which there is a separate repository. Binary releases will then no longer be published for TensorFlow Lite.
Further information on functional changes and Bux fixes can be found in the blog and the release notes.
(who)
