It is very much a crisis and one that academia has no interest to solve.
In a sane world we'd start with figuring out what the mystery meat in the version of pytorch used for training is and work up from there. There is zero funding for building up the required infrastructure for doing actual science instead of whatever the fuck you call the current mess.
Basically anything on a more complex dataset than cifar is probably a statistical fluke because of hyper parameter tuning. To find this out you will need to spend three months building up all the bit rotten code that doesn't work from the paper. To draw the conclusion that everyone is full of shit you'll need to waste years of your life.