I understand. > you can't simply upgrade the software on all the machines You can update a website. And a website can use the local gpu. The break the traini

I understand. > you can't simply upgrade the software on all the machines You can update a website. And a website can use the local gpu. The break the training task in smaller pieces, parallelize and distribute it part is indeed beyond my skills.