What it is
PyTorch is an open-source Python library for building and training neural networks, with GPU-accelerated tensor operations and automatic differentiation. It also ships a C++ distribution (LibTorch) and Java bindings, selectable from the same install page. The project is governed by the PyTorch Foundation, a vendor-neutral body hosted by the Linux Foundation since September 2022; its governing board includes AMD, AWS, Google Cloud, Meta, Microsoft Azure and Nvidia, and the Foundation also stewards related projects such as vLLM, DeepSpeed, Ray, Helion and Safetensors.
Interface and dependencies
Install with pip3 install torch torchvision torchaudio (choosing a CPU-only or CUDA 11.8/12.6/12.8 build); the pytorch.org selector currently offers Pip, LibTorch or Source only — there is no official Conda channel, though conda-forge community builds remain available for those who want conda. The current stable release requires Python 3.10 or later. Training code is standard Python plus torch.nn, torch.optim and the torch.distributed backend for multi-GPU or multi-node runs. TorchScript is deprecated as of PyTorch 2.10 in favor of torch.export, and TorchServe's GitHub repo was archived in August 2025 and is now in limited maintenance with no further updates or security patches. Official Deep Learning Containers and AMIs are published for AWS, Google Cloud and Azure.
Real use and limits
A developer training a computer-vision or NLP model writes it directly in Python against torch.nn, debugs in eager mode with normal Python tooling, then exports via torch.export (the current recommended path now that TorchScript is deprecated) or self-hosts serving, keeping in mind that TorchServe is archived and no longer actively maintained. GPU builds are tied to a specific CUDA or ROCm version, so the install command has to be reselected per target machine rather than being a single universal package. There is no managed inference endpoint from the PyTorch Foundation itself: running models at scale means self-hosting or going through a cloud partner such as AWS SageMaker or Azure Machine Learning.
