PyTorch 1.12: TorchArrow, Functional API for Modules and nvFuser, are now available https://pytorch.org/blog/pytorch-1.12-released/ This explains new features much better: TorchArrow - but why reinvent Pandas? What is the logic? Can someone explain? To build more abstractions around data? But Why? Functional API for Modules - seems like a niche cool feature, but why, again? Complex32 and Complex Convolutions in PyTorch - seems cool for scientific computing, but niche? Forward-mode Automatic Differentiation - not sure what is the purpose. Can someone explain? Datapipes - not sure why they are needed at all. Maybe to unify enterprise benchmarking? Idk again. functorch - looks like again I am out of target audience. The audience is scientific computing, right? All performance improvements are cool, but they are weirdly tied to particular hardware product lines. Are they planning to crystallize the ML only for the selected few?