Libtorch Binaries

This repository provides tooling for compiling libtorch binaries.

Binaries

libtorch	aarch64 Linux	arm64 macOS
2.0.1	download	download
1.13.0	download	download
1.12.0	download	download
1.11.0	download	download
1.10.0	download	download
1.9.0	download	-
1.8.0	download	-
1.7.0	download	-

Usage

Image generation

Cross-compilation in this project is conducted within Docker containers to simplify the sand-boxing of system libraries, such as GLIBC. dockerfiles contains Dockerfiles catered to various operating systems (and the associated compilers that they typically ship with.) To generate the images, ensure your Docker daemon is running and execute

./main.sh dockerbuild

docker compose build --parallel

to build the images following the recipes from the docker-compose.yaml file. For reference, the current images ordered by operating system and version are:

Operating System	GLIBCXX	GLIBC	Python	Docker image tag
Ubuntu 22.04	3.4.30	2.35	3.10	libtorch-factory:Ubuntu22.04
Ubuntu 20.04	3.4.28	2.31	3.8	libtorch-factory:Ubuntu20.04
Ubuntu 18.04	3.4.25	2.27	3.8	libtorch-factory:Ubuntu18.04

If you have a need for a particular operating system that is not listed, or a specific version of GLIB, please open an issue!

To determine the version of GLIBCXX used by your system, use something like the following (where aarch64-linux-gnu is replaced depending on the machine.)

strings /usr/lib/aarch64-linux-gnu/libstdc++.so.6 | grep GLIBCXX

For x86 Linux one would use:

strings /usr/lib/x86_64-linux-gnu/libstdc++.so.6 | grep GLIBCXX

To determine the version of GLIBC used by your system, use ldd --version.

Library compilation

Once at least one docker image has been generated, libtorch may be compiled within the container. It's worth noting that you will need to run your image on the target CPU architecture on which you intend to compile your final product. For instance, if you are deploying code to an ARM platform, you will need to compile libtorch from a similar ARM platform (not an x86_64 platform.)

To compile libtorch, simply run the following subbing <IMAGE> for one of the Docker image tags in the above table.

./main.sh <IMAGE> build

For instance, to compile on Ubuntu22.04, one would use

./main.sh libtorch-factory:Ubuntu22.04 build

The build routine will dump the C++ library and a Python wheel to the local build directory.

To build for a specific version of PyTorch, one can use:

./main.sh <IMAGE> "PYTORCH_VERSION=<VERSION> ./build.sh"

For instance, to compile 1.13.0 on Ubuntu22.04, one would use:

./main.sh libtorch-factory:Ubuntu22.04 "PYTORCH_VERSION=1.13.0 ./build.sh"

If more exotic modifications need to be made, one can open an interactive shell session within the container using:

./main.sh <IMAGE> bash

Compilation outside of a container

If you would prefer to directly compile libtorch outside of a Docker container, this can be accomplished using:

./build.sh

The build routine will dump the C++ library and a Python wheel to the local build directory.

N.B.: Because the build script will alter your python environment, it is recommended to first create a virtual environment and source the terminal before launching the build script.

Packaging this repository

If you need to zip this repository up to move it to a machine, this can be accomplished using:

./main.sh package

Kautenja/libtorch-binaries