Presumably that it was easier to use to illustrate the actual content as opposed to implementation details. This is definitely the case, at least in TensorFlow

Presumably that it was easier to use to illustrate the actual content as opposed to implementation details. This is definitely the case, at least in TensorFlow 1.x there were a lot of details which made converting from [the way we think about the problem] to [practical implementation] more cumbersome. Some of these existed for good reason (eg. performance) while others were simply architectural cruft which the benefit of hindsight makes unnecessary.