Finally, while TensorFlow is the reason we built this infrastructure, its algorithms are independent of TensorFlow itself: the same compiler transformation can extract any computation that executes asynchronously from the host program while communicating through sends and receives. This is useful and can be applied to anything that represents computation as a graph, including other ML frameworks, other kinds of accelerators (for cryptography, graphics, transcoding, etc), and general distributed systems programming models based on graph abstractions. We are interested in exploring new applications of this algorithm in the future.
https://github.com/tensorflow/swift/blob/master/docs/DesignOverview.md#graph-program-extraction 