TensorFlow-DirectML Roadmap - September 2021

Public Release of TensorFlow-DirectML 1.15.5

TensorFlow-DirectML 1.15.5 marks our first non-preview release of TensorFlow 1.x with DirectML support! We've spent over a year improving the performance and stability of the DirectML fork of TensorFlow, and we feel it's ready to be widely used in many scenarios without the preview tag.

Versioning for TensorFlow-DirectML 1.x

Each preview release of TensorFlow-DirectML tracked a base version of upstream TensorFlow. For example, our final preview build (TensorFlow-DirectML 1.15.5.dev210429) was based on top of all the changes in upstream TensorFlow 1.15.5. The .devYYMMDD pre-release tag we appended to the base version allowed us to continue making changes without bumping the base version; we had two preview builds based on upstream 1.15.5, and we had three preview builds based on 1.15.3.

With the release of TensorFlow-DirectML 1.15.5 we can no longer use the pre-release tag to differentiate our builds, so our versions will no longer be in sync with upstream TensorFlow. We intend to continue the current semantic versioning in any future builds of TFDML 1.x, so future releases would be 1.15.6, 15.7, and so on. Performance updates (i.e. new versions of DirectML) will likely be considered patch updates (i.e. updating the last digit only) since they do not introduce new features.

Note that upstream TensorFlow will not be releasing patches anymore; TF 1.15.5 is expected to be the final build of Google's official TensorFlow repository, so there should be no overlapping versions of TFDML patch builds.

Development Moving Forward

The public release of TensorFlow-DirectML 1.15.5 also signals a shift in our team's development focus: we will continue to support the TFDML 1.x builds with bug fixes and performance improvements (through DirectML redistributable updates), but we currently plan keep the overall feature set, like kernels, stable with what we've already shipped. New feature work will be focused on TF 2.x in the interest of catching up to the upstream release schedule and integrating the DirectML backend with the latest features.

We expect that much of the engineering effort that has gone into TFDML 1.15 (mainly kernels) will be applicable in newer versions of TF as well. Stay tuned for more updates on TF 2.x soon!