Is there an existing issue for this?

• I have searched the existing issues

Bug description

When specifying gputouse (at least) in predict_videos.analyze_videos(), it has no effect and it's always GPU 0 that gets used (see image below).

Now, I found this online, while led me to try hardcoding the CUDA_VISIBLE_DEVICES environment variable before relevant imports (tensorflow, most likely), and in this case the specified GPU is effectively used.

Same exact code (.py script) used both times, trying to get CPU 2 engaged.
Also, I used to have CUDA 11.8 installed, as well as previous versions of NVIDIA drivers and of DLC, and it used to work fine.
Lastly, I have CUDA_MODULE_LOADING=LAZY in my environment variables.

Operating System

Ubuntu 22.04

DeepLabCut version

DLC 2.3.5

DeepLabCut mode

single animal

Device type

NVIDIA RTX A4500
NVIDIA drivers 535.54.03
CUDA 12.2

Steps To Reproduce

No response

Relevant log output

No response

Anything else?

No response

Code of Conduct

• I agree to follow this project's Code of Conduct