Pytorch is one of the most popular machine learning frameworks, and its latest iteration (PyTorch 2.0) landed just a couple of days back. Among other things, PyTorch 2.0 offers faster performance with a fully backward-compatible API that guarantees the development ergonomics that PyTorch is known for.
In this talk, we will examine how practitioners (researchers and engineers) can benefit from optimizations provided by PyTorch 2.0 and what other improvements are on the horizon.
In this talk, we will talk about the following:
(i) Performance benefits provided by PyTorch 2.0, even without changing even a single line of code. (ii) How to leverage the best of PyTorch 2.0, sometimes by changing just a single line of code. (iii) A quick overview of "behind-the-scenes" technologies underpinning PyTorch 2.0 - TorchDynamo, AOTAutograd, PrimTorch, and TorchInductor. (iv) API improvements to PyTorch 2.0 that makes the user experience even more ergonomic.