Windows WSL / Docker Desktop users must update to Docker Desktop v4.31.1

Windows WSL users will fail execution on the GPUs from the containers if they are running Docker Desktop versions older than 4.31.1 . My project was running a PyTorch image because it already had cupy in it.

I found this after installing the latest NVIDIA workstation but probably isn’t directly related to that. This seems like some kind of driver issue outside Workbench.

The error message I saw when using cupy was

"CUDARuntimeError: cudaErrorSymbolNotFound: named symbol not found"

video driver 555.xx and later


PS C:\Users\joe> nvidia-smi
Wed Jul 10 23:26:27 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 555.85                 Driver Version: 555.85         CUDA Version: 12.5     |

See

Please tick the appropriate box to help us categorize your post
Bug or Error
Feature Request
Documentation Issue
Other

Your guess is correct. The latest 555.x driver branch on Windows includes new features that require an update to NVIDIA Container Toolkit.

The NVIDIA Container Toolkit installation is managed by Docker Desktop when using Docker Desktop with WSL to run containers. In Docker Desktop 4.31.0 they upgraded to CTK 1.15.0.

Simply update Docker Desktop if you are in this situation and everything should start working again.

Should there be a new version of workbench that verifies the docker version?

The latest installer can install as a non working environment. At least the update from a couple days ago

1 Like

Does Podman require an update also on Windows?

You don’t need to update podman. This was an issue with Docker.

The podman and container toolkit installs are managed by Workbench within the WSL2 distribution, so this problem doesn’t arise.

1 Like