Skip to content

Comment on Infrastructure setup and open-source scripts to train 70B model from bare metal

Comments

I wonder if it's possible for a huge number of hobbyists to team up and train a model together in a distributed manner like seti@home or folding@home. Or does this kind of workload not really lend itself to that approach?

Those things were of course characterised by the ability to spread the work into pretty self-contained work packages. Not sure if that can be done with model training.

Not likely to work. Not many (any) hobbyists can get 400gbps network throughput between each other's GPUs...

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.