Stable Diffusion not working!

Posted by ABOhiccups

Stable Diffusion was installed with NVIDIA GPU. When I tried to open Stable Diffusion. It ask me to setup Tailscale or public domain that are not really required for docker container that does not use secured password. I entered http://localhost:7860 as shown and nothing. After setting up public domain. It still brings up the same setup domain window. I enter Stable Diffusion through a domain name. The page loads as normal. When I'm trying to generate an image. I get an error. RuntimeError: CUDA error: no kernel image is available for execution on the device CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDALAUNCHBLOCKING=1 Compile with TORCHUSECUDADSA to enable device-side assertions. I'm using NVIDIA Pro 4000 SFF Graphics Card on MS-A2 with AMD Ryzen 9. I am running on latest version of Spark Box (Experimental Version) v1.7.37

3 replies

Chris wrote:

Sorry this has been frustrating. There are two separate things here. The setup window: the Stable Diffusion page has no login of its own, so SparkBox only opens it through the Access picker on its card, either a private Tailscale link or a domain you put in front of it. localhost:7860 only works on the machine SparkBox is running on, so if you typed it from another computer it will show nothing. If the domain you set up is reachable from the internet, anyone who finds it can use the page, so a Tailscale link is the safer choice. The CUDA error: "no kernel image is available for execution on the device" means the graphics card was reached, but the software inside the image was not built for that card's generation. This is our current best explanation, not a confirmed cause. One read-only check will tell us which card it is. On the machine running SparkBox, run: nvidia-smi --query-gpu=name,computecap --format=csv Post the output (the card name and the number), and tell us whether SparkBox runs on plain Linux, Windows with WSL, or a NAS. Nothing on your box needs to change for this check.

ABOhiccups wrote:

name, computecap NVIDIA RTX PRO 4000 Blackwell SFF Edition, 12.0

Chris wrote:

Thanks, that settles the card. An NVIDIA RTX PRO 4000 Blackwell SFF Edition reports compute capability 12.0. That means the card itself is being reached, but the Stable Diffusion image SparkBox runs appears not to include code built for that generation, which is why generation stops with this error. Our earlier reply called this our best explanation; with your card name it is now the likely cause, though we have not yet confirmed which part of the image is missing it. Nothing you have set up is wrong. Reinstalling or changing settings will not clear this error, so there is no need to reset anything. Until the image is updated to one that supports this card, images will not generate on the GPU. Stable Diffusion still runs on the CPU, just much more slowly. We have logged this against Stable Diffusion so the team can work on it.