Alessio Fanelli: belief

Transcript context

…Yeah, that's wonderful. Well, we'll talk about the challenges as well that open source has, open models versus closed models. Maybe you want to touch a little bit on PyTorch before we move on to the sort of Meta AI in general. Yeah, we kind of touched on PyTorch in a lot of episodes. So we had George Hotz from TinyGrad. He called PyTorch a CISC and TinyGrad a RISC. I would love to get your thoughts on PyTorch design direction as far as, I know you talk a lot about kind of having a happy path to start with and then making complexity hidden away but then available to the end user. One of the things that George mentioned is I think you have like 250 primitive operators in PyTorch, I think TinyGrad is four. So how do you think about some of the learnings that maybe he's going to run into that you already had in the past seven, eight years almost of running PyTorch? Yeah, I think there's different models here, but I think it's two different models that people generally start with. Either they go like, I have a grand vision and I'm going to build a giant system that achieves this grand vision and maybe one is super feature complete or whatever. Or other people say they will get incrementally ambitious, right? And they say, oh, we'll start with something simple and then we'll slowly layer out complexity in a way that optimally applies Huffman coding or whatever. Like where the density of users are and what they're using, I would want to keep it in the easy, happy path and where the more niche advanced use cases, I'll still want people to try them, but they need to take additional frictional steps. George, I think just like we started with PyTorch, George started with the incrementally ambitious thing. I remember TinyGrad used to be, like we would be limited to a thousand lines of code and I think now it's at 5,000. So I think there is no real magic to which why PyTorch has the kind of complexity. I think it's probably partly necessitated and partly because we built with the technology available under us at that time, PyTorch is like 190,000 lines of code or something at this point. I think if you had to rewrite it, we would probably think about ways to rewrite it in a vastly simplified way for sure. But a lot of that complexity comes from the fact that in a very simple, explainable way, you have memory hierarchies. You have CPU has three levels of caches and then you have DRAM and SSD and then you have network. Similarly, GPU has several levels of memory and then you have different levels of network hierarchies, NVLink plus InfiniBand or Rocky or something like that, right? And the way the flops are available on your hardware, they are available in a certain way and your computation is in a certain way and you have to retrofit your computation onto both the memory hierarchy and like the flops available. When you're doing this, it is actually a fairly hard mathematical problem to do this setup, like you find the optimal thing. And finding the optimal thing is, what is optimal depends on the input variables themselves. So like, okay, what is the shape of your input tensors and what is the operation you're trying to do and various things like that. Finding that optimal configuration and writing it down in code is not the same for every input configuration you have. Like for example, just as the shape of the tensors change, let's say you have three input tensors into a Sparstar product or something like that. The shape of each of these input tensors will vastly change how you do this optimally placing this operation onto the hardware in a way that will get you maximal throughput. So a lot of our complexity comes from writing out hundreds of configurations for each single PyTorch operator and templatizing these things and symbolically generating the final CUDA code or CPU code.…

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