TensorFlow 2.18 works with NumPy 2 and Hermetic CUDA

The light version 2.18 of TensorFlow now has NumPy 2.0 installed and compiled by default. Now use a hermetic CUDA and create new GPUs with a rating (compute capacity) of 8.9 on Nvidia grid.

Anzeige

When installing NumPy, there are 2 functions that cause the TensorFlow Team the most API problems, if problems occur with fehlermeldungen, conversions out of bounds or a large number of scalar representations can be performed. The team has the NumPy-1-Verhalten available for the Python conversion. Ferner sollten Entwicklerinnen und Entwickler the new Rules of Type Promotion of NumPy 2 beachten, die auch Auswirkungen auf TensorFlow-Ergebnisse and are Fehlern or unpräzizen Ergebnissen of Skalaren führen führen. Support for NumPy 1.26 will now strengthen the team until 2025.

CUDA is no longer local

Entwicklerinnen and Entwickler, the TensorFlow from the Quellen construction, have a hermetic CUDA. TensorFlow is connected to CUDNN and NCCL to Netz, soft drink CUDA is no longer installed locally. As Grund, the TensorFlow Team is reproducing builds for Google’s own ML projects. Develop the team TensorRT in CUDA-Builds with “code health improvement”.

TensorFlow Binaries now provide kernels for GPUs with a compute capacity (compute capacity) of 8.9 for Ada GPUs, which includes new hardware with NVIDIA RTX 40**, L4 or L40 found. The Pascal-Generation (6.0) GPUs are no longer available. The team has used useful GPUs in version 2.16 to run TensorFlow itself on its platform.

Once the TensorFlow Lite team on RTLite has put an end to the years, the contributors can use their own Repository account. For TensorFlow Lite, the binary releases became more light.

You can find more information about features and fixes on the blog and release notes.

(WHO)