#Machine Learning

PyTorch 2.4 was released on July 24, 2024, adding torch.compile support for Python 3.12, AOTInductor freezing for CPU, a new libuv-based TCPStore backend that speeds up large-scale job startup, and an easier Python custom operator API. Here is what changed.

PyTorch 2.1 was released on October 4, 2023, adding automatic dynamic shape support in torch.compile, distributed checkpointing across multiple ranks, torch.compile support for the NumPy API, and a prototype torch.export mechanism. Here is what changed.

PyTorch 2.0 was released on March 15, 2023, a next-generation release that introduced torch.compile as the main API, supercharging performance at the compiler level while keeping the same eager-mode experience and full backward compatibility. Here is what changed.