Wing Lian: evaluation

Transcript context

…Cool. The last few things I wanted to cover was also just the more advanced stuff that you covered yesterday. So I'll just, caution this as like, yeah, this is more advanced. But you mentioned Stackllama and Multipack. What are they and what should people know? Yeah, so, so, Stack Llama was, that paper came out, so Stack Llama I think was like, two, two, two separate, two separate concepts that they announced, so the first one was They being hugging face. Yeah, sorry, yes, they being hugging face, so the first one being sort of like, this idea of packing, like some packing sequences together, so like, if we think about training data, right, your training data is, let's say, to keep the math easy, let's say your training data is 500, We, we, we, we will use the terminology words. Let's say your training data is 500 words long, and let's say your, your context length, you know how much data your, that your model can accept is like, or that you want feed into your model. It's, let's say, we won't use tokens again, we'll we'll use it is it's 4,000 tokens, right? So if you're training at 4K Con or four 4,000 4K contacts and you're only using 500 of it, you're sitting like with the other 1500. 3, 500 words that you're not using, right? And typically that's either filled with these PAD tokens, so I think I made the analogy last night that it's like having sort of like a glass here you fill it up with a shot of liquor and then you're and that's your training data and then you just fill it up with more water and those are your PAD tokens and it's just, it doesn't do much, right? It's still the same thing, but you still have to go through all of that to go through all your training data. And then, so what Stack Llama showed was you could just sort of take your training data, append the next row of training data until you filled that entire 4k context, so in this example, right, with 500 words to 4k, that's 8 rows of training data. But, the problem with that is, is that with a lot of these transformer models, they're very much relying on attention, right? So, like, if you now have this sequence of words that now, in order for the, the model has seen all of these other words before, right? And then it sees another set of words, another set of words, but it's learning everything in context of all the words that it's seen before. We haven't corrected the attention for that. And just real quickly, since I said that that paper was two concepts, the other one was, I believe it was like a reinforcement learning, but outside the scope of this. So going from that, I implemented that early on because I was like, Oh, wow, this is really great. And. Yes, because it saves you a bunch of time, but the trade off is a little bit of accuracy, ultimately, but it still did pretty well. I think when I did Manicore, I think it used sort of that concept from Stack Llama of just sort of appending these sequences together, right? And then sort of the next evolution of that is Multipack, right? k when I did Manicore, I think it used sort of that concept from Stack Llama of just sort of appending these sequences together, right? And then sort of the next evolution of that is Multipack, right? So, there was a separate paper on that, it was, I believe it was referenced, it got referenced in the Orca paper, where you could, you could properly mask those out using like a, I think it was like a lower block triangular attention mask, and then sort of, so, So, there's that. I did try implementing that, manually recreating that mask, but then one from the OpenChat, so he was helping with OpenOrca as well, and he had done an implementation of Multipack, and where he used FlashAttention, so FlashAttention So that was released by TreeDAO, and it was this huge performance gain. Everybody uses it now, even the Transformers library now, they've taken all of these, like, people are taking all of these models and sort of like, making it compatible with FlashAttention. But in Flash Tension, there is one particular implementation that lets you say, Well, I'm sending you all of these sequences like you would in Stack Llama, But let me send you another, another, Set of information about, this is where this set of sequences is, this is where the second set of sequences is. So like, if it was like, 500 words long, and you stacked them all together, you would just send it a row of information that was like, 0, 500, 1000, 1500, etc, etc, out to 4000. And it would know, alright, I need to break this up, and then run the forward pass with it. And then it would be able to, and it was much more, much more performant. And I think you end up seeing like 10x, 20x improvements over sort of, I mean, I think FlashAttention was like a 2x improvement, and then adding that with the Multipack, you start to see like, depending on, how much data you have, up to like a 20x improvement sometimes. 20x. 20x. Wow. Yeah. And I only know the 20x because I, like, before last night, I was like, I re ran the alpaca, I looked up the alpaca paper because it was like, I just need a frame of reference where somebody did it, and I think they used eight A100s for three hours, and they said it cost them 100. I don't, I don't think eight A100s cost, I don't know how much it costs right now. But I ended up rerunning it. Usually a dollar an hour, right? Yeah, so eight. The cheapest is like a…

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