Soumith Chintala: preference

Transcript context

…That's interesting. And we kind of touched on this with Chris Lattner when he was on the podcast. If you think about frameworks, they have the model target. They have the hardware target. They have different things to think about. He mentioned when he was at Google, TensorFlow trying to be optimized to make TPUs go brr, you know, and go as fast. I think George is trying to make especially AMD stack be better than ROCm. How come PyTorch has been such as Switzerland versus just making Meta hardware go brr? First, Meta is not in the business of selling hardware. Meta is not in the business of cloud compute. The way Meta thinks about funding PyTorch is we're funding it because it's net good for Meta to fund PyTorch because PyTorch has become a standard and a big open source project. And generally it gives us a timeline edge. It gives us leverage and all that within our own work. So why is PyTorch more of a Switzerland rather than being opinionated? I think the way we think about it is not in terms of Switzerland or not. We actually the way we articulate it to all hardware vendors and software vendors and all who come to us being we want to build a backend in core for PyTorch and ship it by default is we just only look at our user side of things. Like if users are using a particular piece of hardware, then we want to support it. We very much don't want to king make the hardware side of things. So as the MacBooks have GPUs and as that stuff started getting increasingly interesting, we pushed Apple to push some engineers and work on the NPS support and we spend significant time from Meta funded engineers on that as well because a lot of people are using the Apple GPUs and there's demand. So we kind of mostly look at it from the demand side. We never look at it from like oh which hardware should we start taking opinions on. Is there a future in which, because Mojo or Modular Mojo is kind of a superset of Python, is there a future in which PyTorch might use Mojo features optionally?…

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