Alessio Fanelli: preference

Transcript context

…How much per call? Free, because it's not commercial. I'm like, hopefully nobody from Entropic is listening. But yeah, it works. But I think that's kind of like the sweet spot. And then the completion length is like 1,800, you know, so it's not like it stays within the 60K band. But anyway, yeah, curious to see. And I think another thing from your Twitter model parades that I really like is actually differentiating between the type of workload. I feel like people talk about these models as like anything you do is like the same thing, but you posted about GPT 3.5 being cheaper than LLAMA 2 for completion-heavy workloads. What does that mean? Yeah, so there are different terms, I guess, based on like whatever community you're in. So I think in the research community, they probably call it pre-filling is handling prompt tokens. And then I believe maybe decoding is what they call generating completion tokens. We'll just use prompt tokens and completion tokens. But for prompt tokens, the work, it's entirely compute bound. And the reason why is the same reason why transformers are so good at being kind of being trained in parallel. And it's that you can parallelize the entire sequence or you can parallelize an input, not just along the batch dimension, but the sequence dimension. So that means let's look at the first layer of the transformer. Imagine like that entire layer could fit in memory. I just read that to memory. And then I basically apply the matrix multiplication of the entire sequence on this layer. If you're doing token generation, instead, you have to read the layer, then taking the first input, and then you have to read the next layer and then do that same input. You have to do it all the way to the end of the model. And then you generate the next token and that next token passes through all the layers again. So before what you were doing is you have all your input tokens in parallel, they're going through the first layer. So you read the first layer, then in parallel, they're going through the second layer. You read the second layer, so on and so forth to the end. But when you're doing it for one token at a time, you read the first layer, second layer, third layer, fourth, blah, blah, blah. Then you do it all over again for the next token. And so as a result, for your sequence length N, you end up using N times more memory bandwidth than compute.…

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