Hi,
My current setup is composed of:
- laptop under Ubuntu 22.10 (Thunderbolt 4)
- eGPU enclosure (Razer Core)
- 4070 TI
My initial setup was running using the ubuntu provided drivers: 525-open
I had to add the nvidia.NVreg_OpenRmEnableUnsupportedGpus=1
kernel parameter.
(it does not work with the 525
, only with the 525-open
)
With this setup:
- the GPU is properly detected:
nvidia-smi
andnvidia-settings
see the card - I can use the GPU under Blender or to run Stable Diffusion inference
The problems began when I tried to use PyTorch and DreamBooth for Stable Diffusion.
My understanding is that to get this to run, I need to upgrade to Cuda 11.6 or 11.7.
Because ubuntu packages only provide Cuda 11.5, I use the installed you provide.
I run the NVidia installer for Cuda-11.7.
It automatically forces me to upgrade to nvidia-driver-530
dkms status
nvidia/530.30.02, 5.19.0-35-generic, x86_64: installed
With the 530 driver, my 4070ti is no longer detected:
nvidia-smi
returns a “No devices were found”nvidia-settings
does not see the card either
I also tried to install cuda 12.1, but ended up with the exact same result.
If I cleanup everything and go back to 525-open and Cuda-11.5 then the GPU is usable again.
What can I do to diagnose the problem and hopefully fix this?
(It seems that Ubuntu 23.04 will have Cuda 11.7 by default, but this is early to switch).
Tiry