[Feature Request] DirectML as Pluggable device for TensorFlow

[HighCWu]

In the latest version of Tensorflow (v2.5.0), PluggableDevice support has been added.
At the beginning, I paid attention to your dml implementation and apple's apple/tensorflow_macos implementation at the same time. Recently I suddenly discovered that their repo has been archived. After checking carefully, I know that they directly use PluggableDevice to make tensorflow run on metal.
This plug-in approach, I think, should also be very suitable for directml.

[PatriceVignola]

Hi @HighCWu

This repository is dedicated to TensorFlow 1.15, which doesn't have PluggableDevice support. A PluggableDevice plugin for TensorFlow 2.x is our very next step though and we already started investigating and working on it, so stay tuned!

[HighCWu]

Great, I am looking forward to it.

[zhmlcg]

I hope I can contribute to the development of tensorflow2-directml so that it will be released sooner. But I would only mess things up thanks to my poor background ;-)

[marunguy]

Is there any progress?

[PatriceVignola]

There's been a lot of progress :)

We cannot say exactly when, but we're planning the release of a preview package very soon. I'll reply to this issue once it's released!

[oscarbg]

+1 to be notified :-)

[value808]

+1 to be notified myself. thanks, this is gold!

[Kwinpeng]

+1 to be notified.

[rjdyk]

DirectML is officially available as a PluggableDevice for TensorFlow🎉!

Our preview package can be found on PyPI and the project is fully open-source.

This is the initial preview of the DML pluggable device so please leave any feedback or suggestions on GitHub.

[HighCWu]

Great, I will try it.