There's a lot of effort involved in the training runs too, and you might be able to get away with saying the ops engineers made creative choices too (of which checkpoints were good etc.)

Since it costs millions to produce one of these models, it's not just taking the software and running it to compile them.