2

I recently set up WSL2 (Windows 10, 21H2) with CUDA/Docker support to train some neural networks. For this I followed the instructions from this guide: https://docs.nvidia.com/cuda/wsl-user-guide/index.html

Once I was done, I ran an image as specified in the document I linked:

docker run -it --gpus all -p 8888:8888 tensorflow/tensorflow:latest-gpu-py3-jupyter

At this point I am able to launch the browser and work on a Jupyter Notebook with GPU support. I verify this by running:

import tensorflow as tf
tf.config.list_physical_devices('GPU')

The output confirms the GPU is available: [PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')].

After I'm done I save my changes and stop CTRL-C out of the container.

Up to this point, everything works as I expect.

To resume work I've been using docker start -i happy_bose where happy_bose is the name of the container I ran earlier.

However, GPU support is non-existent (the call to tf.config.list_physical_devices('GPU') returns an empty list).

I was expecting docker start -i happy_bose to start the container with GPU support but this is not the case. Is there any way to reuse a container with GPU support or do I need to create it and start it using docker run every time?

Adeiln
  • 131
  • 4
  • What version of Windows are you running? If you are not already running 21H2 you should try upgrading to that version before you do anything. – Ramhound Jan 05 '22 at 15:04
  • I am on 21H2 (updated the question to mention this fact). That being said, the problem is not that it does not work at all but that it does not work when I restart a previous container. Unlike `docker run`, `docker start` does not take a `--gpus all` argument, but I think this should not matter if the container was first run using said flag. – Adeiln Jan 05 '22 at 18:29

0 Answers0