MPS & Deep learning libraries

When MPS is enable on GPU system does anything need to be changed the way we use Tensorflow (or any other deep learning libraries we use) or it is transparent to the applications written using these libraries?

Any disadvantages in enabling MPS?


I think this question is still pretty relevant.

I’ve found the related issue on Tensorflow’s github — which states that MPS doesn’t support CUDA “stream callbacks” (you can find that remark in official MPS documentation), but Tensorflow uses them a lot.

On the other side, Caffe / IBMCaffe is known to work flawlessly with MPS (and even gets some boost from it).

Any plans from NVIDIA team to support Tensorflow (esp. stream callbacks) for MPS?