Libtorch

libtorch is the C++ distribution of PyTorch, providing a native C++ interface to the PyTorch framework. It exposes the same core capabilities as the Python API—tensor computation, automatic differentiation, and neural network building—but targeted at C++ applications and deployment environments. libtorch is distributed as a binary library that can be linked into C++ projects, enabling developers to build and run PyTorch models without Python.

Core components include the ATen tensor library for tensor operations, the autograd system for automatic differentiation, and a neural network API (the nn module) for defining and training models. The TorchScript component offers serialization and deployment of models; the JIT compiler can optimize and run scripted or traced models in production. CUDA support allows using NVIDIA GPUs with cuDNN-accelerated kernels. libtorch also exposes a subset of PyTorch’s distributed and parallel utilities for multi-GPU training where available.

Typical usage involves compiling with CMake or Bazel and linking against the libtorch binaries, for example via CMake's find_package(Torch) and target_link_libraries. Prebuilt binaries are provided for Linux, Windows, and macOS, and require a compatible C++14 (or newer) toolchain and compatible CUDA toolkit if GPU support is needed.

Licensing for libtorch follows the PyTorch project’s BSD-3-Clause license, with some components under third‑party licenses. Libtorch is maintained by the PyTorch community, originally developed by Facebook AI Research and now supported by contributors worldwide.