Episode 33: Tri Dao, Stanford: On FlashAttention and sparsity, quantization, and efficient inference
Generally Intelligent
00:00
The Butterfly Factorization, Nicely Modeled 2019
The idea is to find a parameterization that can contain a lot of these fast transforms. And then you can learn over the set of parameters. The problem we ran into was that there is this kind of combinatorial aspect of searching over different algorithms. For example, at each stack, you might choose to combine input element number zero with input element number seven or eight or nine or ten. There are so many choices that you have to make.
Transcript
Play full episode
