machinelearning Higher CUDA / Tensorflow compatibility

Is your feature request related to a problem? Please describe. The currently used libraries like CUDA and Tensorflow are ancient and have not been updated since 2019.

Describe the solution you'd like I would like to use the latest or newer libraries like CUDA, and TensorFlow to use their new abilities.

Describe alternatives you've considered Or share simple a roadmap of ML.Net, because currently, it looks already dead.

Jul 15 '24 11:07 sascha-michalke

it's not dead but seems that is not to far from there

Nov 14 '24 12:11 rblanca

@luisquintanilla are you still taking PRs against this issue? We've had to update ML.Net to use Tensorflow 2.7 to support Ampere cards natively and happy to push up the code.

Reason for this being that when the image classification models trained with Tensorflow 2.3.2 are transpiled from native Turing to Ampere architecture the models load but give frozen results. i.e. no matter what image is supplied the output is the same. Running CPU only results are fine. Card drivers stripped back to the metal and rebuilt but problem persisted so we cracked open the code.

May 01 '25 15:05 RossHNPC

So, the only solution would be a rebuild of all old models? I would be totally fine with that solution, instead of having a dead platform.

Oct 14 '25 09:10 sascha-michalke