Back in the elden days, I took a course called "Large Scale Scientific Computing". It was mostly about multiplying large matrices. I didn't think this was going to be remotely applicable to anything commercial.
Only read the first section but this sounds really impressive -- up to 50% of up to 17% of training time when using the Muon optimiser, so up to around 7% of basically pure improvement with no downside.
Ah ... the temptation of the optimizer. It's such a simple algorithm, it has far more impact on back-propagation calculations than ... the actual backprop calculation, never mind details like model architecture. So tempting to work on it.
But so very, very, very, very hard to make progress on it. Even at PhD level. Just don't try ...
Back in the elden days, I took a course called "Large Scale Scientific Computing". It was mostly about multiplying large matrices. I didn't think this was going to be remotely applicable to anything commercial.
Boy was I wrong.
A superior alternative to standard Muon and AdamW optimizers for training large models.
Fantastic work, instantly valuable, immediately usable.
A big THANK YOU to the authors:
Jack Zhang, Noah Amsel, Berlin Chen, and Tri Dao
Tri Dao's lab must have saved countless watts with FlashAttention. Great to see them continuing to open-source massive efficiency gains.
Only read the first section but this sounds really impressive -- up to 50% of up to 17% of training time when using the Muon optimiser, so up to around 7% of basically pure improvement with no downside.
Ah ... the temptation of the optimizer. It's such a simple algorithm, it has far more impact on back-propagation calculations than ... the actual backprop calculation, never mind details like model architecture. So tempting to work on it.
But so very, very, very, very hard to make progress on it. Even at PhD level. Just don't try ...