Building for Windows on ARM64

1. Prerequisites
2. Cloning PyTorch
3. Preparing Environment
4. Setting Environment Variables
5. Build
   1. PyTorch
   2. LibTorch

Prerequisites

1. Build Tools for Visual Studio 2022 OR any version of Visual Studio 2022

   • Under Workloads, select Desktop development with C++

   • Under Individual Components, select ARM64/ARM64EC build tools (latest)

2. Python 3.12 Windows installer (ARM64)

3. Download and install Arm Performance Libraries for Windows

   • %ARMPL_DIR% environment variable should be automatically added to your PATH by default installation settings.

4. Install Rust

5. Building libuv (v1.39.0) from source (we'll use the folder path in Build)

   • Open ARM64 Native Tools Command Prompt for VS2022 from start menu
   • Change path to the suitable folder
   • Run following command to build libuv

   git clone https://github.com/libuv/libuv.git -b v1.39.0
   
   cd libuv
   
   echo Clean folder...
   git clean -fxd
   
   echo Configuring libuv...
   mkdir libuv\build
   cd libuv\build
   cmake .. -DBUILD_TESTING=OFF
   
   echo Building libuv...
   cmake --build . --config Release
   
   echo Installing libuv...
   cmake --install . --prefix ../install
   
   pause
   

6. (OPTIONAL) You may install sccache for speeding up future builds by cache support.

   • You need to add sccache to your PATH manually.

Cloning PyTorch

1. Clone the PyTorch repository (or your fork)

   git clone --recursive https://github.com/pytorch/pytorch
   cd pytorch
   

2. Init and sync submodules

   git submodule sync
   git submodule update --init --recursive
   

Preparing Environment

NOTE: Since currently conda doesn't support Windows ARM64, instead we're recommending to use built-in Python venv functionality for similar isolating experience. (more info)

1. Creating a virtual environment (venv)

   python -m pip install --upgrade pip
   python -m venv .venv
   echo * > .venv\.gitignore
   call .\.venv\Scripts\activate
   

2. Install requirements

   pip install -r requirements.txt
   

3. Activate Arm64 Native Tools Command Prompt for VS 2022 environment by calling one of the following commands based on what you've installed.

   • For Build Tools:

     "C:\Program Files (x86)\Microsoft Visual Studio\2022\BuildTools\VC\Auxiliary\Build\vcvarsall.bat" arm64

   • For Visual Studio 2022 Community:

     "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Auxiliary\Build\vcvarsall.bat" arm64

   • For Visual Studio 2022 Enterprise:

     "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" arm64

   • For Visual Studio 2022 Preview:

     "C:\Program Files\Microsoft Visual Studio\2022\Preview\VC\Auxiliary\Build\vcvarsarm64.bat" arm64

Setting Environment Variables

1. Set the following environment variables:

   set BLAS=APL
   set USE_LAPACK=1
   set MSSdk=1
   set DISTUTILS_USE_SDK=1
   :: please change {YOUR_LIBUV_FOLDER} to the your libuv folder
   set libuv_ROOT={YOUR_LIBUV_FOLDER}\install
   

2. Copy libuv dll to the lib folder to include it for packaging (please check: Prerequisites)

   copy %libuv_ROOT%\lib\Release\uv.dll torch\lib\uv.dll
   

3. (OPTIONAL) If you want to use sccache for speeding up your future build by cache also set following.

   set CMAKE_C_COMPILER_LAUNCHER=sccache
   set CMAKE_CXX_COMPILER_LAUNCHER=sccache
   

Build

You can generate a .whl (wheel) file for PyTorch (the Python package) OR a .zip file for LibTorch (the C++ core). If you'd like to use PyTorch right away without creating a wheel file, you can build it directly in your environment. Here are your options:

PyTorch

1. OPTION 1 - Generating and Installing Wheel File (for distribution or installation on other systems):

   Call following command to start build and generate output

   :: set PYTORCH_BUILD_VERSION=2.6.0   ## OPTIONAL: If you wish to give a specific versioning for `.whl` file 
   :: set PYTORCH_BUILD_NUMBER=1        ## (for instance you can specify 2.6.0 as shown) 
   python setup.py bdist_wheel
   

   Output file will be under dist folder with .whl extension.

   Format will look like torch-FOO+gitBAR-cp312-cp312-win_arm64.whl OR if you specify version, format will be like torch-2.6.0-cp312-cp312-win_arm64.whl

   Now you can copy and install this .whl file anywhere in your system by calling

   pip install torch-YOURVERSION-cp312-cp312-win_arm64.whl
   

2. OPTION 2 - Directly Installing to Local Environment (for development on your system only)

   Run one of the following commands to initiate the build:

   • Install the package in development mode, ideal when making frequent changes to codebase

   python setup.py develop 
   

   • Install the package for regular use, rather than ongoing development

   python setup.py install 
   

LibTorch

1. Additional preparation for libtorch folder

   :: Options: Release, Debug or RelWithDebInfo
   set CMAKE_BUILD_TYPE=Release
   
   :: Prepare the environment
   mkdir libtorch
   mkdir libtorch\bin
   mkdir libtorch\cmake
   mkdir libtorch\include
   mkdir libtorch\lib
   mkdir libtorch\share
   mkdir libtorch\test
   

2. Call following command to start build and generate output

   python ./tools/build_libtorch.py
   

3. Prepare .zip file from libtorch output

   :: Moving the files to the correct location
   move /Y torch\bin\*.* libtorch\bin\
   move /Y torch\cmake\*.* libtorch\cmake\
   robocopy /move /e torch\include\ libtorch\include\
   move /Y torch\lib\*.* libtorch\lib\
   robocopy /move /e torch\share\ libtorch\share\
   move /Y torch\test\*.* libtorch\test\
   move /Y libtorch\bin\*.dll libtorch\lib\
   
   :: Create output under dist
   mkdir dist
   tar -cvaf dist/libtorch-win.zip -C libtorch *
   
   :: Cleanup raw data to save space
   rmdir /s /q libtorch
   

4. Now you can copy and extract dist/libtorch-win.zip file anywhere in your system

   • For C++ example, you can visit Installing C++ Distributions of PyTorch

   • You can just add following line to your CMakeLists.txt file.

     set(CMAKE_PREFIX_PATH "C:/YOUR_CORRESPONDING_PATH/libtorch-win")