Tri Dao: prediction

Transcript context

…Yeah, the I.O. awareness is probably one of the biggest innovations here. And the idea behind it is, like you mentioned, the FLOPS growth of the cards have been going up, but the memory bandwidth, not as much. So I think maybe that was one of the assumptions that the original attention paper had. So talk a bit about how that came to be as an idea. It's one of those things that like in insight, it's like, obviously, why are we like rewriting to like HBM every time, you know, and like once you change it, it's clear. But what was that discovery process? Yeah, in hindsight, a lot of the ideas have already been there in the literature. And I would say is it was somehow at the intersection of both machine learning and systems. And you kind of needed ideas from both sides. So on one hand, on the system side, so lots of systems folks have known that, oh, you know, kernel fusion is great. Kernel fusion just means that instead of performing, you know, loading the same element, instead of performing an operation, write it down, load it back up and perform the second operation, you just load it once, perform two operations and then write it down again. So that saves you kind of memory read and write in the middle there. So kernel fusion has been a classic. There's been other techniques from the system side, like tiling, where you perform things in the form of computations in block, again, so that you can load it into a really fast memory. Think of it as a cache. And this is, again, classical computer science ideas, right? You want to use the cache. So the system folks have been thinking about these ideas for a long time, and they apply to attention as well. But there were certain things in attention that made it difficult to do a complete kernel fusion. One of which is there is this softmax operation in the middle, which requires you to essentially sum across the row of the attention matrix. So it makes it difficult to kind of break it, because there's this dependency. So it makes it difficult to break things into a block. So on the system side, people have been thinking about these ideas, but it's been difficult to kind of do kernel fusion for the entire operation. On the machine learning side, people have been thinking more algorithmically. They say, okay, either we can approximate attention, or there's this trick called the online softmax trick, which says that because of softmax, the way it's written mathematically, you can actually break it up into smaller pieces, do some rescaling, and still get the right answer. So this online softmax trick has been around for a while. I think there was a paper from NVIDIA folks back in 2018 about this. And then there was a paper from Google. So Marcus, Rob, and Stats wrote a paper late 2021 on using this online softmax trick to break attention up into smaller pieces. So a lot of the ideas were already there. But it turns out, you kind of need to combine ideas from both sides. So you need to understand that, hey, we want to do kernel fusion to reduce memory written writes. But we also need this online softmax trick to be able to break the softmax into smaller pieces so that a lot of the systems tricks kind of carry through. We saw that, and it was kind of a natural idea that we ended up using ideas from both sides, and it ended up working pretty well. Yeah. Are there any downsides to kernel fusion? If I think about databases and the reasons why we have atomic operations, you know, it's like, you have observability and fallback in between them. How does that work with attention? Is there anything that we lose by fusing the operations?…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.