Install pytorch with Homebrew, apt, dnf

Project links

• Homepage
• Repository
• Package manager page

Commands and aliases

• torchrun

Project history

Development began in 2016 among contributors from the Lua Torch community, with substantial support from Facebook AI Research and participation from other organizations. PyTorch launched publicly in 2017 as a tensor and automatic-differentiation framework designed for flexible neural-network research.

The PyTorch 1.0 effort combined PyTorch's research-oriented frontend with production capabilities associated with Caffe2 and ONNX. Later compiler work added `torch.compile` while retaining eager execution as the normal programming model.

In 2022, governance moved from Meta to the PyTorch Foundation under the Linux Foundation. This established a neutral organizational home supported by major cloud and accelerator companies while technical direction remained with project maintainers.

Adoption history

PyTorch initially won adoption among AI researchers through its imperative, Python-first interface and dynamic computation graphs. The project's first-year retrospective documented a rapidly growing ecosystem of domain libraries, research use, meetups, and model-export work.

PyTorch 1.0 broadened that appeal toward production by incorporating Caffe2-derived deployment capabilities and introducing a path from eager research code to tracing, compilation, and export. The project subsequently became shared infrastructure across academia, cloud providers, hardware vendors, and AI application developers.

How it is used

Researchers and developers use tensors, autograd, `torch.nn`, optimizers, data utilities, and domain libraries to build and train models on CPUs, GPUs, and other accelerators.

Infrastructure teams use distributed primitives and `torchrun` for multi-process training. Deployment workflows may use compilation, export, quantization, native C++ integration, or specialized runtimes from the wider PyTorch ecosystem.

Why package nerds care

PyTorch helped make eager tensor programs written as ordinary Python a mainstream deep-learning workflow. Package maintainers care about its unusually broad binary surface: Python bindings, native libraries, CPU and accelerator backends, distributed launch tools, and closely coordinated domain packages must all remain compatible.

The Homebrew formula is useful where PyTorch is managed as system software rather than solely inside a Python environment. Its `torchrun` executable also makes the package directly relevant to command-line and distributed-training workflows.

Timeline

• 2016: Development begins among members of the Lua Torch community.
• 2017: PyTorch launches publicly.
• 2018: The PyTorch 1.0 initiative combines the research frontend with Caffe2 production capabilities.
• 2022: PyTorch joins the Linux Foundation under the new PyTorch Foundation.
• 2023: PyTorch 2.0 introduces `torch.compile` while preserving eager-mode development.

Related projects

• Lua Torch supplied the project lineage and part of its original contributor community.
• Caffe2 contributed production-oriented capabilities during the development of PyTorch 1.0.
• ONNX provided an interchange route for exporting models to other runtimes and hardware stacks.
• The wider official ecosystem includes torchvision, torchaudio, TorchRL, ExecuTorch, torchao, and distributed and compiler components.

Sources

• https://github.com/pytorch/pytorch#readme
• https://pytorch.org/blog/PyTorchfoundation/
• https://pytorch.org/blog/a-year-in/
• https://pytorch.org/blog/pytorch-the-open-language-of-ai/
• https://pytorch.org/blog/the-road-to-1_0/