I'm opening this PR for @Bogdan-Wiederspan who did the actual code prototyping and implementation in this PR. However, we needed access to the CI pipelines, which is why the feature branch is not part of his fork but already in the main repo.
This PR adds all tools necessary
to inspect graphs regarding their AOT/XLA compatibility,
to specialize signatures of SavedModels (reading a graph with batching support from a SavedModel and converting it to one or more graphs that each have a fixed batch size), and
to AOT compile this specialized graphs.
New unit tests are provided along with docstrings for all new helpers.
Besides these additions, a few minor changes to existing code are provided, mostly to align the existing tensorflow tools to python 3.
I'm opening this PR for @Bogdan-Wiederspan who did the actual code prototyping and implementation in this PR. However, we needed access to the CI pipelines, which is why the feature branch is not part of his fork but already in the main repo.
This PR adds all tools necessary
New unit tests are provided along with docstrings for all new helpers.
Besides these additions, a few minor changes to existing code are provided, mostly to align the existing tensorflow tools to python 3.
Once merged, we should
@valsdav