Build and Install PyTorch🔗
Official instructions for building and installing PyTorch can be found here.
In the sections below we provide a short version that can be used to build and install PyTorch C++ API files (LibTorch) for use with vAccel. We assume that the required dependencies are already installed.
Build PyTorch C++ API files (LibTorch)🔗
Clone the PyTorch repo, adjusting TORCH_VERSION to the desired version:
TORCH_VERSION=2.6.0
git clone https://github.com/pytorch/pytorch --depth 1 --recursive \
-b "${TORCH_VERSION}"
cd pytorch
Build the source code:
export _GLIBCXX_USE_CXX11_ABI=1
# If CUDA dependencies are installed set this to `1` for CUDA support
export USE_CUDA=0
cmake -S . -B build
cmake --build build --parallel "$(nproc)"
and install the binaries:
[Alternative] Install pre-built PyTorch C++ API files (LibTorch)🔗
PyTorch provides pre-built binaries for LibTorch. Ie. for version 2.6.0:
Download and extract CPU-only binaries:
wget https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-2.6.0%2Bcpu.zip
unzip libtorch-cxx11-abi-shared-with-deps-2.6.0+cpu.zip
or download and extract binaries with CUDA support (here for CUDA 11.8):
wget https://download.pytorch.org/libtorch/cu118/libtorch-cxx11-abi-shared-with-deps-2.6.0%2Bcu118.zip
unzip libtorch-cxx11-abi-shared-with-deps-2.6.0+cu118.zip
and move files to the desired installation directory: