PyTorch was primarily developed by which company?

Facebook, now Meta. PyTorch was built at Facebook AI Research (FAIR) and released as open source in January 2017. It is a Python library for deep learning that provides tensors with GPU acceleration and automatic differentiation.

It grew out of Torch, a Lua-based library FAIR already used, and its original authors include Adam Paszke, Sam Gross, Soumith Chintala and Gregory Chanan. Its defining feature, inspired by Chainer, is define-by-run: the computation graph is built as ordinary Python code executes, so standard debuggers and control flow just work. Facebook merged its Caffe2 framework into PyTorch in 2018, and in September 2022 Meta moved the project to the PyTorch Foundation under the Linux Foundation.

Interviewers often contrast it with TensorFlow, released by Google in 2015, whose 1.x versions made you define a static graph first; TensorFlow 2.0 switched to eager execution by default in 2019. PyTorch 2.0 (2023) added torch.compile to get compiled-graph speedups from eager code. Do not confuse PyTorch with the older Lua Torch, which is no longer developed.