1 comment

[ 2.2 ms ] story [ 15.4 ms ] thread
I'm not 100% sure that this buys you a ton since the multinode training is still done with tensorflow clusters or horovod, so there's nothing new I can see (after all of 10 minutes of looking, so take that with a grain of salt) with the actual distributed pieces. That said, workflow management is still an important piece of the training puzzle. I've seen a lot of good quantitative work put to poor use for lack of a good workflow tool.