Skip to content

PyTorch 2 Internals [slides]

drive.google.com
2 pointsperone1 comment
On HN

Comments

As an end user, who cares? Make it more like Keras. Quit messing about w tensor optimization because you'll get smoked by someone else's performance eventually.

Last time I checked, there wasn't even a function to shuffle a tensor. Something you'd do at the start of every training iteration.

AboutSource Built by g1lg1l

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