I was looking into getting Stable Diffusion running privately to simplify my blog workflow while still keeping my data in-house. My primary desktop is a 3rd generation i5, so it's practically worthless for this. The only hardware I have available to me for this project is a 3rd-gen i7 with an NVIDIA Tesla K20Xm running Proxmox 8.3 server.
By default, Stable Diffusion only works with CUDA 3.7 and above. If this project is successful, we will be able to run Stable Diffusion on an NVIDIA graphics card that only supports CUDA 3.5. Even using older graphics hardware is faster than CPU-only image generation.
I know you can passthrough the GPU and use it, I've done that before.
Steps
These steps should be turned into future posts.
- Install the latest NVIDIA Data Center Driver 470.256.02 for the Telsa K20Xm on the Proxmox host using the non-free drivers and in the container using the runfile.
- Side quest: Create systemd unit files that limit power use on startup
- Create an (un)privileged container and passthrough the GPU in a multi-container friendly way using snippets.
- Install the same NVIDIA Data Center Driver installed on the host inside the container and install the latest compatible CUDA Toolkit 11.4 using the runfile.
- Do not install the driver from the CUDA Toolkit, it's older than the one we installed.
- Explore the ideas presented in Running on Kepler #13286 on the AUTOMATIC1111/stable-diffusion-webui GitHub page.
Update
Unfortunately if it's possible, I just do not have the skills to pull this off.
