CUDA driver fails to initialize in WSL

[ivangplus]

Is there an existing issue for this?

• I have searched the existing issues

Current Behavior

When running profiler in WSL for CUDA-enabled Pytorch code, CUDA driver fails to initialize. Specifically, this happens after entering the memray.Tracker context manager.

$ python -m memray run -q -c 'import torch; print(torch.cuda.is_available())'
False

Expected Behavior

CUDA driver being able to initialize,

$ python -m memray run -q -c 'import torch; print(torch.cuda.is_available())'
False

Steps To Reproduce

1. WSL2 (Ubuntu 24.04), CUDA 12.7
2. Prepare python env:

   $ pip install memray
   $ pip install torch --index-url https://download.pytorch.org/whl/cu126

3. Run python -m memray run -q -c 'import torch; print(torch.cuda.is_available())'

Memray Version

1.16.0

Python Version

3.12

Operating System

Linux

Anything else?

I checked that this happens exactly in WSL environment. If I try to reproduce this in a Docker container of ubuntu:22.04 image, Docker engine being installed in the same WSL environment, CUDA successfully initializes.
Having read about supported environments, I'm still not quite sure if WSL is supported or not.