The point was that the concurrency model of these ML libraries is not dependent on a CPU for backprop (ideally), so discussing language level concurrency as an advantage is irrelevant. The parallelism is not implemented at the language level, nor does it rely on language constructs for synchronizing shared memory. The parallelism happens at the hardware level (e.g. with SIMD). And in the model architecture (e.g. with convolutions).

So the advantages you posited in your lengthy diatribe are meaningless, because you failed to take into account the problem domain, along with utterly failing to grasp what gets calculated during backprop. Plus, backprop doesn't get calculated in Python.

The argument about functional vs imperative languages is old, tired and is not relevant to the subject. Please at least learn some fundamentals about NNs and ML frameworks before contributing a long, uninformed response.