Is there a way to pass {'device_id': 0} to DmlExecution Provider ?

[alanoo81]

Windows on an AMD laptop here with discrete GPU.
Default DirectML seems to use the iGPU

Got the stuff working, but it seems to only use the GPU 1 which in this case is the iGPU (Radeon 680M, so quite weak for those tasks)

I can see reference of options for DmlExecutionProvider, but haven't found a way to pass the device_id to it

I can understand we should so something like this : w-okada/voice-changer#410
But can't figure out myself with my low skills.

Can someone help ?

[AVGRadmin]

Are you running the source code or the pre-built version?
If source code:
You could add os.environ["CUDA_VISIBLE_DEVICES"] = "0" to force cuda to use the First GPU(0) by hiding the second one.

os.environ["CUDA_VISIBLE_DEVICES"] = "0" = Use the first.
os.environ["CUDA_VISIBLE_DEVICES"] = "1" = use the second (and so on).

This would have to be added in the core.py at the very top somewhere.

Alternatively you could try to set the environment arg BEFORE starting the app. I'm a pure linux user tho, so i'm not 100% certain that this would solve it, but you could try. I'm also just assuming that the Pre-Built app will use this cuda os enviroment arg, but i see no reason why not.
You might have to start the app (bat/exe) from the same CLI window as the one you pass CUDA_VISIBLE_DEVICES=X to.
Worst case, reboot your PC to discard/reset CUDA_VISIBLE_DEVICES env.

Mr.GPT as a reference since i'm not entirely sure about windows:

I'm using this method myself to select which GPU to use, but i'm on linux and doing it in my docker-compose file.

[alanoo81]

I'm running the python source version, but not CUDA unfortunately, DirectML because AMD hardware, so CUDA env variables would not work

[AVGRadmin]

Oh yeah that's right. My bad. I'm not used to AMD personally myself and forgot about that tiny important fact.. facepalm xD