Installing LibTorch

Download and install the PyTorch C++ distribution for your platform

Overview

LibTorch is the C++ distribution of PyTorch. It contains all headers, libraries, and CMake configuration files required to depend on PyTorch from C++.

System Requirements

Before installing LibTorch, ensure your system meets these requirements:

Linux

• GLIBC: Version 2.29 or newer (for cxx11 ABI version)
• GCC: Version 9 or newer (for cxx11)
• CMake: Version 3.18 or higher
• Disk Space: At least 2GB for LibTorch binaries

Windows

• Visual Studio: 2017, 2019, or 2022
• CMake: Version 3.18 or higher
• Disk Space: At least 2GB for LibTorch binaries

macOS

• macOS: 10.13 (High Sierra) or later
• Xcode: Latest version recommended
• CMake: Version 3.18 or higher

Download LibTorch

Binary Distribution

Download pre-built LibTorch from the PyTorch website.

1. Select Your Configuration

   Choose your platform, compute platform (CPU/CUDA), and language (C++).

2. Download the Archive

   Download the appropriate LibTorch distribution for your system.

   CPU-only (Linux):

   wget https://download.pytorch.org/libtorch/nightly/cpu/libtorch-shared-with-deps-latest.zip
   unzip libtorch-shared-with-deps-latest.zip
   

   GPU-enabled (Linux with CUDA 12.4):

   wget https://download.pytorch.org/libtorch/nightly/cu124/libtorch-shared-with-deps-latest.zip
   unzip libtorch-shared-with-deps-latest.zip
   
3. Note the Installation Path

   Remember the absolute path to the extracted libtorch directory - you’ll need this for CMake configuration.

For GPU support, select the LibTorch distribution matching your CUDA version. Find the right link using the selector on pytorch.org.

From PyTorch Python Installation

If you have PyTorch installed via pip, you can query the LibTorch path:

python3 -c 'import torch;print(torch.utils.cmake_prefix_path)'

Building from Source

Prerequisites

To build PyTorch from source, you need:

• Python: 3.10 or later
• Compiler: Full C++20 support (GCC 11.3.0+ on Linux, clang on macOS)
• Disk Space: At least 10GB free
• Build Time: 30-60 minutes for initial build

Clone the Repository

git clone https://github.com/pytorch/pytorch
cd pytorch
git submodule sync
git submodule update --init --recursive

Install Dependencies

Linux

# From the PyTorch directory
pip install --group dev
pip install mkl-static mkl-include

# For CUDA support: Add LAPACK support for GPU
.ci/docker/common/install_magma_conda.sh 12.4

Build PyTorch

Linux

export CMAKE_PREFIX_PATH="${CONDA_PREFIX:-'$(dirname $(which conda))/../'}:${CMAKE_PREFIX_PATH}"
python -m pip install --no-build-isolation -v -e .

For CUDA support on Linux, ensure CUDA is properly installed and USE_CUDA=1 is set (default). To disable CUDA, set USE_CUDA=0.

Creating Your First Application

Project Structure

Create a minimal LibTorch application with this structure:

example-app/
  CMakeLists.txt
  example-app.cpp

CMakeLists.txt

Create a CMakeLists.txt file:

cmake_minimum_required(VERSION 3.18 FATAL_ERROR)
project(example-app)

find_package(Torch REQUIRED)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${TORCH_CXX_FLAGS}")

add_executable(example-app example-app.cpp)
target_link_libraries(example-app "${TORCH_LIBRARIES}")
set_property(TARGET example-app PROPERTY CXX_STANDARD 17)

# Windows: Copy DLLs to avoid memory errors
if (MSVC)
  file(GLOB TORCH_DLLS "${TORCH_INSTALL_PREFIX}/lib/*.dll")
  add_custom_command(TARGET example-app
                     POST_BUILD
                     COMMAND ${CMAKE_COMMAND} -E copy_if_different
                     ${TORCH_DLLS}
                     $<TARGET_FILE_DIR:example-app>)
endif (MSVC)

Application Code

Create example-app.cpp:

#include <torch/torch.h>
#include <iostream>

int main() {
  torch::Tensor tensor = torch::rand({2, 3});
  std::cout << tensor << std::endl;
  return 0;
}

Including <torch/torch.h> is the most reliable way to access the PyTorch C++ API. For more fine-grained control, you can include specific headers.

Build the Application

1. Create Build Directory

   mkdir build
   cd build
   
2. Configure with CMake

   Specify the path to your LibTorch installation:

   cmake -DCMAKE_PREFIX_PATH=/absolute/path/to/libtorch ..
   

   Or if using PyTorch from pip:

   cmake -DCMAKE_PREFIX_PATH=`python3 -c 'import torch;print(torch.utils.cmake_prefix_path)'` ..
   
3. Build the Project

   cmake --build . --config Release
   
4. Run the Application

   ./example-app
   

   Expected output (values will vary due to randomness):

   0.2063  0.6593  0.0866
   0.0796  0.5841  0.1569
   [ CPUFloatType{2,3} ]
   

On Windows, debug and release builds are not ABI-compatible. If building in debug mode, use the debug version of LibTorch. Always specify the correct configuration in cmake --build.

Platform-Specific Notes

Visual Studio Extension (Windows)

For Windows developers not using CMake, the LibTorch Project Template Visual Studio extension can automatically configure project settings and link options.

Features:

• Automatic setup for debug and release configurations
• Pre-configured include paths and library links
• Easy project creation

See the demo video for usage instructions.

CUDA Support

For GPU-enabled builds:

1. Install NVIDIA CUDA: Download from NVIDIA
2. Install cuDNN: Version 8.5 or above from NVIDIA
3. Compatible Compiler: Check CUDA compatibility

Set the CUDA path if installed in a non-standard location:

export PATH=/usr/local/cuda-12.4/bin:$PATH

ROCm Support (AMD GPUs)

For AMD GPU support on Linux:

1. Install ROCm: Version 4.0+ from AMD
2. Set ROCm Path (if not in /opt/rocm):
   export ROCM_PATH=/path/to/rocm
   

Troubleshooting

CMake Cannot Find Torch

Problem: CMake Error: Could not find a package configuration file provided by "Torch"

Solution: Ensure CMAKE_PREFIX_PATH points to the absolute path of your LibTorch installation:

cmake -DCMAKE_PREFIX_PATH=/absolute/path/to/libtorch ..

Linking Errors on Windows

Problem: Unresolved external symbols or DLL not found errors

Solution:

• Verify debug/release configuration matches LibTorch build
• Ensure DLL copy step in CMakeLists.txt is present
• Check that Visual Studio version is compatible

Missing OpenMP (macOS)

Problem: OpenMP-related errors during build

Solution: Install OpenMP via Homebrew:

brew install libomp

Getting Help

If you encounter installation issues:

• Forum: PyTorch Discussion Forum
• GitHub: PyTorch Issues
• Documentation: Official C++ Docs

Next Steps

Now that you have LibTorch installed, learn about:

• Tensor Operations - Working with tensors in C++
• Autograd - Automatic differentiation
• Modules - Building neural networks