Install Windows preview#

First, you must enable “Optional diagnostic data”, otherwise Windows cannot join Windows Insiders.

Then, join Windows Insiders program with Microsoft account (an account can be created if you don’t have one: https://insider.windows.com/). The channel must be Dev:

Then run Windows Update. It will download the pre=release installer. Windows will restart and reinstall.

Install WSL 2 preview#

In Windows, make sure the following Windows features are enabled:

• WSL: dism.exe /online /enable-feature /featurename /all /norestart
• Virtual machine platform: dism.exe /online /enable-feature /featurename /all /norestart

Now restart Windows, then Windows will have WSL and the wsl command line tool. Run Windows Update again to get the latest WSL 2. When this is done, in the update history, it must show 4.19.121 or later:

Then manually install this patch: https://wslstorestorage.blob.core.windows.net/wslblob/wsl_update_x64.msi. And then run the following command as administrator to update the kernel to the latest version:

C:\WINDOWS\system32>wsl —update Checking for updates… Downloading updates… Installing updates… This change will take effect on the next full restart of WSL. To force a restart, please run ‘wsl —shutdown’. Kernel version: 5.4.72

Now go to Microsoft Store, install a Linux distribution. I installed the first one Ubuntu, it is the same as the third item Ubuntu 20.04 LTS

When the installation is done, use the following command to set the default version to 2 and verify:

C:\WINDOWS\system32>wsl —set-default-version 2 For information on key differences with WSL 2 please visit https://aka.ms/wsl2 The operation completed successfully.

C:\WINDOWS\system32>wsl —list —verbose NAME STATE VERSION * Ubuntu Stopped 2 docker-desktop Stopped 2 docker-desktop-data Stopped 2

WSL version can also be set for a specific distribution:

C:\WINDOWS\system32>wsl —set-version ubuntu 2 Conversion in progress, this may take a few minutes… For information on key differences with WSL 2 please visit https://aka.ms/wsl2 The distribution is already the requested version.

Now launch the installed Ubuntu (default) from the Store, or using the command wsl:

For the first time launch, it asks for initializing the user name and password.

Install Nvidia driver preview and CUDA toolkit#

In Windows, download the preview driver from Nvidia: https://developer.nvidia.com/cuda/wsl/download and install.

Then launch WSL, install CUDA:

apt-key adv —fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub sh -c ‘echo “deb http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64 /” > /etc/apt/sources.list.d/cuda.list’ apt update apt install -y cuda-toolkit-11-1

Install Docker and Nvidia container toolkit#

First, if Docker Desktop is installed in Windows, turn off the WSL integration for WSL in the distro, because It does not work with CUDA in WSL:

Now go to WSL, install Docker from there:

curl https://get.docker.com | sh

Then install Nvidia container toolkit in WSL:

distribution=$(. /etc/os-release;echo$ID$VERSION\_ID) curl -s -L [https://nvidia.github.io/nvidia-docker/gpgkey](https://nvidia.github.io/nvidia-docker/gpgkey) | sudo apt-key add - curl -s -L [https://nvidia.github.io/nvidia-docker/](https://nvidia.github.io/nvidia-docker/)$distribution/nvidia-docker.list | sudo tee /etc/apt/sources.list.d/nvidia-docker.list curl -s -L https://nvidia.github.io/libnvidia-container/experimental/$distribution/libnvidia-container-experimental.list | sudo tee /etc/apt/sources.list.d/libnvidia-container-experimental.list sudo apt update sudo apt install -y nvidia-docker2

When it is done, restart docker in WSL:

sudo service docker stop sudo service docker start

Run WSL + CUDA + Docker + Jupyter + TensorFlow#

Finally, Jupyter notebook can be used in WSL to run popular AI framework code, like TensorFlow. Nvidia’s official tutorial runs TensorFlow with Jupyter notebook in Docker image tensorflow/tensorflow. However, this docker image is no longer updated and gives errors. Now there is another up-to-date image tensorflow/tensorflow. In WSL, run :

docker run -it —gpus all -p 8889:8888 tensorflow/tensorflow

Here container’s port 8888 is mapped to WSL’s port 8889, because on Windows I am already running an instance of Jupyter notebook, which already occupies port 8888.

Dock shows the following messages:

________ _______________ ___ __/__________________________________ ____/__ /________ __ __ / _ _ \_ __ \_ ___/ __ \_ ___/_ /_ __ /_ __ \_ | /| / / _ / / __/ / / /(__ )/ /_/ / / _ __/ _ / / /_/ /_ |/ |/ / /_/ \___//_/ /_//____/ \____//_/ /_/ /_/ \____/____/|__/

WARNING: You are running this container as root, which can cause new files in mounted volumes to be created as the root user on your host machine.

To avoid this, run the container by specifying your user’s userid:

$docker run -u$(id -u):$(id -g) args…

[I 17:15:59.736 NotebookApp] Writing notebook server cookie secret to /root/.local/share/jupyter/runtime/notebook_cookie_secret jupyter_http_over_ws extension initialized. Listening on /http_over_websocket [I 17:15:59.947 NotebookApp] Serving notebooks from local directory: /tf [I 17:15:59.948 NotebookApp] The Jupyter Notebook is running at: [I 17:15:59.948 NotebookApp] http://ccd3c60790ab:8888/?token=8b1969bc9a278498fd5debe119feb6d86130850166425bef [I 17:15:59.948 NotebookApp] or http://127.0.0.1:8888/?token=8b1969bc9a278498fd5debe119feb6d86130850166425bef [I 17:15:59.948 NotebookApp] Use Control-C to stop this server and shut down all kernels (twice to skip confirmation). [C 17:15:59.953 NotebookApp]

To access the notebook, open this file in a browser: file:///root/.local/share/jupyter/runtime/nbserver-1-open.html Or copy and paste one of these URLs: http://ccd3c60790ab:8888/?token=8b1969bc9a278498fd5debe119feb6d86130850166425bef or http://127.0.0.1:8888/?token=8b1969bc9a278498fd5debe119feb6d86130850166425bef

Now take the last URI, replace “127.0.0.1:8888” with “localhost:8889”, go to Windows and launch browser, paste the URI: http://localhost:8889/?token=8b1969bc9a278498fd5debe119feb6d86130850166425bef, The browser should load the Jupyter notebook with built-in TensorFlow tutorial:

However, running the cells in the notebook, it shows some errors. If it shows an error of git not found (“FileNotFoundError: No such file or directory: ‘git’”), then run these commands in a cell to install git:

!apt update !apt install -y git

If the python code throws an error “InternalError: Blas GEMM launch failed” or “InternalError: Blas GEMV launch failed”, the fix is to run these code:

#import tensorflow as tf physical_devices = tf.config.list_physical_devices(‘GPU’) tf.config.experimental.set_memory_growth(physical_devices[0], True)

The it should work:

Encoding and decoding video with GPU in WSL?#

Not supported yet. To test GPU encoding and decoding, the latest version of FFmpeg can be installed:

sudo add-apt-repository ppa/graphics sudo add-apt-repository ppa/multimedia sudo add-apt-repository ppa/ffmpeg4 sudo apt-get update sudo apt-get upgrade && sudo apt-get dist-upgrade sudo apt-get install ffmpeg

Then install the Nvidia encoding/decoding libraries:

sudo apt install libnvidia-decode-455 sudo apt install libnvidia-encode-455

Now try to use FFmpeg to decode a video and encode it with hevc_nvenc:

ffmpeg -hwaccel auto -i input.mkv -map 0:v:0 -map 0 -map_metadata 0 -loglevel verbose -c hevc_nvenc -profile main10 -pix_fmt yuv420p10le -preset slow -tune film -b_adapt 2 -b_ref_mode middle -bf 3 -rc vbr_hq -deblock 0:0 -rc-lookahead 25 -rc_lookahead 25 -g 250 -keyint_min 23 -refs 4 -sc_threshold 40 -temporal_aq 1 -spatial_aq 1 -nonref_p 1 -c aac -ar 48000 -b 256k -ac 6 -b 2048k output.nvenc.mp4 –y

It does not work and gives the following error for decoding the input video:

[h264 @ 0x55c4517dd540] decoder->cvdl->cuvidGetDecoderCaps(&caps) failed -> CUDA_ERROR_INVALID_DEVICE: invalid device ordinal [h264 @ 0x55c4517dd540] Failed setup for format cuda: hwaccel initialisation returned error.

And it gives the following error for encoding the output video:

[hevc_nvenc @ 0x55c451065b00] OpenEncodeSessionEx failed: unsupported device (2): (no details) [hevc_nvenc @ 0x55c451065b00] Nvenc unloaded Error initializing output stream 0:0 — Error while opening encoder for output stream #0:0 - maybe incorrect parameters such as bit_rate, rate, width or height

I searched around and found a video of WSLConf’s CUDA session, where Nvidia conforms GPU encoding/decoding is not yet supported in WSL, and will come in the future: