20 releases (breaking)

Uses new Rust 2024

0.17.0	Apr 23, 2025
0.16.0	Jan 14, 2025
0.15.0	Oct 28, 2024
0.13.2	May 3, 2024
0.3.0	Nov 20, 2022

#531 in Machine learning

5,347 downloads per month
Used in 29 crates (5 directly)

MIT/Apache

1.5MB
30K SLoC

Burn Torch Backend

Burn Torch backend

This crate provides a Torch backend for Burn utilizing the tch-rs crate, which offers a Rust interface to the PyTorch C++ API.

The backend supports CPU (multithreaded), CUDA (multiple GPUs), and MPS devices (MacOS).

Installation

tch-rs requires the C++ PyTorch library (LibTorch) to be available on your system.

By default, the CPU distribution is installed for LibTorch v2.6.0 as required by tch-rs.

CUDA

To install the latest compatible CUDA distribution, set the TORCH_CUDA_VERSION environment variable before the tch-rs dependency is retrieved with cargo.

export TORCH_CUDA_VERSION=cu124

On Windows:

$Env:TORCH_CUDA_VERSION = "cu124"

Note: tch doesn't expose the downloaded libtorch directory on Windows when using the automatic download feature, so the torch_cuda.dll cannot be detected properly during build. In this case, you can set the LIBTORCH environment variable to point to the libtorch/ folder in torch-sys OUT_DIR (or move the downloaded lib to a different folder and point to it).

For example, running the validation sample for the first time could be done with the following commands:

export TORCH_CUDA_VERSION=cu124
cargo run --bin cuda --release

Important: make sure your driver version is compatible with the selected CUDA version. A CUDA Toolkit installation is not required since LibTorch ships with the appropriate CUDA runtimes. Having the latest driver version is recommended, but you can always take a look at the toolkit driver version table or minimum required driver version (limited feature-set, might not work with all operations).

Once your installation is complete, you should be able to build/run your project. You can also validate your installation by running the appropriate cpu, cuda or mps sample as below.

cargo run --bin cpu --release
cargo run --bin cuda --release
cargo run --bin mps --release

Note: no MPS distribution is available for automatic download at this time, please check out the manual instructions.

Manual Download

To install tch-rs with a different LibTorch distribution, you will have to manually download the desired LibTorch distribution. The instructions are detailed in the sections below for each platform.

Compute Platform	CPU	GPU	Linux	MacOS	Windows	Android	iOS	WASM
CPU	Yes	No	Yes	Yes	Yes	Yes	Yes	No
CUDA	Yes ^[1]	Yes	Yes	No	Yes	No	No	No
Metal (MPS)	No	Yes	No	Yes	No	No	No	No
Vulkan	Yes	Yes	Yes	Yes	Yes	Yes	No	No

^{The LibTorch CUDA distribution also comes with CPU support.}

CPU

🐧 Linux

First, download the LibTorch CPU distribution.

wget -O libtorch.zip https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-2.6.0%2Bcpu.zip
unzip libtorch.zip

Then, point to that installation using the LIBTORCH and LD_LIBRARY_PATH environment variables before building burn-tch or a crate which depends on it.

export LIBTORCH=/absolute/path/to/libtorch/
export LD_LIBRARY_PATH=/absolute/path/to/libtorch/lib:$LD_LIBRARY_PATH

🍎 Mac

First, download the LibTorch CPU distribution.

wget -O libtorch.zip https://download.pytorch.org/libtorch/cpu/libtorch-macos-x86_64-2.6.0.zip
unzip libtorch.zip

Then, point to that installation using the LIBTORCH and DYLD_LIBRARY_PATH environment variables before building burn-tch or a crate which depends on it.

export LIBTORCH=/absolute/path/to/libtorch/
export DYLD_LIBRARY_PATH=/absolute/path/to/libtorch/lib:$DYLD_LIBRARY_PATH

🪟 Windows

First, download the LibTorch CPU distribution.

wget https://download.pytorch.org/libtorch/cpu/libtorch-win-shared-with-deps-2.6.0%2Bcpu.zip -OutFile libtorch.zip
Expand-Archive libtorch.zip

Then, set the LIBTORCH environment variable and append the library to your path as with the PowerShell commands below before building burn-tch or a crate which depends on it.

$Env:LIBTORCH = "/absolute/path/to/libtorch/"
$Env:Path += ";/absolute/path/to/libtorch/"

CUDA

LibTorch 2.6.0 currently includes binary distributions with CUDA 11.8, 12.4 or 12.6 runtimes. The manual installation instructions are detailed below for CUDA 12.6, but can be applied to the other CUDA versions by replacing cu126 with the corresponding version string (e.g., cu118 or cu124).

🐧 Linux

First, download the LibTorch CUDA 12.6 distribution.

wget -O libtorch.zip https://download.pytorch.org/libtorch/cu126/libtorch-cxx11-abi-shared-with-deps-2.6.0%2Bcu126.zip
unzip libtorch.zip

Then, point to that installation using the LIBTORCH and LD_LIBRARY_PATH environment variables before building burn-tch or a crate which depends on it.

export LIBTORCH=/absolute/path/to/libtorch/
export LD_LIBRARY_PATH=/absolute/path/to/libtorch/lib:$LD_LIBRARY_PATH

Note: make sure your CUDA installation is in your PATH and LD_LIBRARY_PATH.

🪟 Windows

First, download the LibTorch CUDA 12.6 distribution.

wget https://download.pytorch.org/libtorch/cu126/libtorch-win-shared-with-deps-2.6.0%2Bcu126.zip -OutFile libtorch.zip
Expand-Archive libtorch.zip

Then, set the LIBTORCH environment variable and append the library to your path as with the PowerShell commands below before building burn-tch or a crate which depends on it.

$Env:LIBTORCH = "/absolute/path/to/libtorch/"
$Env:Path += ";/absolute/path/to/libtorch/"

Metal (MPS)

There is no official LibTorch distribution with MPS support at this time, so the easiest alternative is to use a PyTorch installation. This requires a Python installation.

Note: MPS acceleration is available on MacOS 12.3+.

pip install torch==2.6.0 numpy==1.26.4 setuptools
export LIBTORCH_USE_PYTORCH=1
export DYLD_LIBRARY_PATH=/path/to/pytorch/lib:$DYLD_LIBRARY_PATH

Note: if venv is used, it should be activated during coding and building, or the compiler may not work properly.

Example Usage

For a simple example, check out any of the test programs in src/bin/. Each program sets the device to use and performs a simple element-wise addition.

For a more complete example using the tch backend, take a loot at the Burn mnist example.

Too many environment variables?

Try .cargo/config.toml (cargo book).

Instead of setting the environments in your shell, you can manually add them to your .cargo/config.toml:

[env]
LD_LIBRARY_PATH = "/absolute/path/to/libtorch/lib"
LIBTORCH = "/absolute/path/to/libtorch/libtorch"

Or use bash commands below:

mkdir .cargo
cat <<EOF > .cargo/config.toml
[env]
LD_LIBRARY_PATH = "/absolute/path/to/libtorch/lib:$LD_LIBRARY_PATH"
LIBTORCH = "/absolute/path/to/libtorch/libtorch"
EOF

This will automatically include the old LD_LIBRARY_PATH value in the new one.

Dependencies

~18MB
~346K SLoC