Skip to content

Comment on Helion: A high-level DSL for performant and portable ML kernels

Comments

Is it normal to spend 10minutes on tuning nowadays? Do we need to spend another 10 minutes upon changing the code?

You mean autotune? I think 10 minutes is pretty normal, torch.compile('max-autotune') can be much slower than that for large models.

Add to that it can be done only once by developers before distribution for major hardware. Configs saved. Then on client side selected.

AboutSource Built by g1lg1l

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