Torch for R on WSL2
R Torch Package
This past spring my laptop of 5 years died and had to be replaced. Unlike the more forward thinking experts out there, I failed to hang on to most any installation scripts, so it’s taken some time to re-build my local work environment to my liking.
I’ve made a commitment to organizing much more on the Windows Subsystem for Linux (Ubuntu 22.04 jammy jellyfish) instead of Windows this time for speed and code interoperability. All but 6 Gb of of the RAM have been allocated to Linux using the .wslconfig trick. But to tap into the full capability of this PC, I want to enable the NVIDIA Gpu-compute for fast deep learning and other data science workloads. Over the past couple of years I have followed the Posit mlverse youtube channel and the development of packages like tabnet and brulee and I’d like to be familiar with what they are capable of.
Daniel Falbel and Javier Luraschi provide a torch installation guide that is a good start, but has a few gaps that I thought I’d write about here, mostly for future me. I’d be thrilled if this blog post happens to help someone else.
WSL installation
Microsoft provides installation instructions. We will assume here that WSL is already fully installed, and that an Ubuntu terminal shell is available.
At the terminal shell, execute
lsb_release -a
and make note of your Ubuntu operating system version.
Video Card and Driver
The first step is checking system information for your machine to confirm that a GPU capable video card is installed. Look for something like NVIDIA GeForce RTX 3070 Ti or higher.
Back at the terminal, execute
nvidia-smi
and make a note of the driver version. Right now, CUDA requires that the driver be at least 530. If not, execute
sudo apt install nvidia-driver-535
or build with an even newer version.
CUDA
I ran the two local WSL-Ubuntu runfile script commands at the terminal. This script makes a few checks for driver versioning and takes care of the installer.
wget https://developer.download.nvidia.com/compute/cuda/12.2.1/local_installers/cuda_12.2.1_535.86.10_linux.run
sudo sh cuda_12.2.1_535.86.10_linux.run
This is a very large package and will require some time to finish.
cuDNN
NVIDIA cudnn has its own installation guide.
In order to download cuDNN, ensure you are registered for the NVIDIA Developer Program.
- Go to: NVIDIA cuDNN home page.
- Click Download.
- Complete the short survey and click Submit.
- Accept the Terms and Conditions. The site provides a list of available cuDNN versions for download.
- Select the cuDNN version that you want to install. I chose cuDNN v8.9.4 for CUDA 12.x.
Move the downloaded deb file from downloads to the Ubuntu /tmp folder.
Skip down to the subsection called “Debian Local Installation” and execute:
sudo dpkg -i /tmp/cudnn-local-repo-ubuntu2204-8.9.4.25_1.0-1_amd64.deb
sudo cp /var/cudnn-local-repo-*/cudnn-local-*-keyring.gpg /usr/share/keyrings/
sudo apt-get update
sudo apt-get install libcudnn8=8.9.4.25-1+cuda12.2
sudo apt-get install libcudnn8-dev=8.9.4.25-1+cuda12.2
in RStudio
If you haven’t already, execute sudo rstudio-server start and launch Rstudio in a browser at localhost:8787
Then inside a console, run
> install.packages(“torch”)
> Sys.setenv(CUDA="11.7")
> torch::install_torch(reinstall = TRUE)
> install.packages("tensorflow")
> tensorflow::install_tensorflow()
Other Resources:
