This includes correct use of `softmax()` which now expects an explicit dimension, and tensor copies in the Exporter.