Hi All..
Total nooob here, am trying to get a couple of models running in comfyui, first one is z-image. Have managed to get ollama and comfyui running, and text generating llm’s are no problem, but trying to get text-image, text-music ect, or anything that requires specific nodes just results in the same error;
CUDA error: no kernel image is available for execution on the device Search for `cudaErrorNoKernelImageForDevice’ in CUDA Runtime API :: CUDA Toolkit Documentation for more information. CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect. For debugging consider passing CUDA_LAUNCH_BLOCKING=1 Compile with `TORCH_USE_CUDA_DSA` to enable device-side assertions.
I have tried various ways of updating comyui, all seemed to work ok, have tried using the template (z-image.json) that came with the model, also tried a template from within comfyui, everything gave the same (above) error.. Any ideas? Searching on the net just ended up with me doing the updates, still have not found a solution..
Am using Linux (debian 13), have 64gb ram and geforce 1070 (8gb).
“Updating” is not always the solution…. you need the correct version of Torch, not necessarily the latest one. For a 1070, I think you need to be running cu126?
Do you run it natively, or in a docker? There are quite a few docker solutions around that might save you lots of hassle.
This one for example: GitHub - YanWenKun/ComfyUI-Docker: 🐳Dockerfile for 🎨ComfyUI. | 容器镜像与启动脚本
Thanks for the responses guys, have given up trying to get anywhere with comfyui several times now, but I keep coming back..
I am (clearly) a noob with all of this, but have been trying to make comfyui work for about three months so far. Every single guide I have followed (hundreds!) has resulted in errors, no exceptions. The endless mis-match of versions on an endless list of software has just become a blur. 99% of my time is spent going round and round the various update functions in comfyui, most of which end up with more errors.
Easy to get text models working, but have not got any other model to work (image/vid etc). Every workflow (have about 20) results in more and different errors. I have spent so many hours chasing down errors that I seem to do little else.
Trying not to whine on about this but I have to say, with all the amazing coding ai models all you guys have, is this really the best you can come up with? I know much of this is in its infancy, but the lack of tools to automate the sorting of all these versions and errors, having to manually chase down all these endless problems is a real pain..
Thanks tanoshimi, how am I supposed to know which version of torch I need? Is there a way to know? Amongst all the errors I am pretty sure I installed cu124 as that was supposed to be the right one acording to a guide I followed. I will do as you suggest, but if there is a way to find out how to do this and find out the right version, I would be grateful to know it.
Hey Arunderan, not running it in docker. I did try that in the early days, could not find a guide that did not result in so many errors that I had to re-install my OS to clear them. Pretty sure I have not tried the one you suggest, will try changing my cu124 to cu126 first, doing it all through docker means starting over again and I have spent a lot of hours downloading endless nodes and models, not keen to do all that again until I have exhausted trying to fix what I have.
Will report back after trying cu126..
just tried to download cu126, it is not available on ;
but there is no file for debian 13 (trixie).
You don’t have to, the docker that i have linked to is quickly set up. And you can change the model paths to the folders you already have.
But yeah, i also had to do some system tweaks over the time now. They had nothing to do with the docker itself though. I was for example trapped for a week with an error regarding ntfs drives. Linux is simply not as easy as Windows. I personally have never managed to get the native version working. That’s where i have given up then ^^
I love Linux, but it does suffer greatly from too many clever coders making poor guides. These folk can be geniuses at understanding all this stuff, but none of that equates to having the skill to put together a decent guide, that is a skill in its own right. Nearly every guide I have seen for anything to do with ai llm models and all the supporting software structures, are full of holes, missing instructions, leap-frogging over essntial steps and a general attitude that us poor noobs should already know how to do do this stuff, or should have to learn it. Thing is, I am a photographer that wants to try and use ai for photo editing, really dont have a spare year or two to learn huge amounts of coding and how to bolt all these bits of software together to make it work. Have been hoping comfyui would be a collection of tools that would manage this lot once installed, but seems it still needs lots of software management and understanding of how these bits work together, and the update system is plain daft, almost feels like a game now, clicking endlessly on all these different update options/restarts/refreshes, bit like a digital whack-a-mole..
That is 4 guides, 3 from nvidia, one off youtube to try and install cuda 126, every one results in ;
sudo sh cuda_13.1.1_590.48.01_linux.run
Installation failed
Will try and track down an error log file, am starting to develop a talent for reading them..
Just that i mention it, Have you considered to buy a cheap Win 11 license? It would save you lots of hassle. At Windows it is really (nearly) as easy as downloading the portable and starting it. The culprit could be your graphics card too by the way. That’s quite an old model. Even when it has enough Vram
From the cuda_installer log;
ERROR]: Install of driver component failed. Consult the driver log at /var/log/nvidia-installer.log for more details.
[ERROR]: Install of 590.48.01 failed, quitting
From the nvidia_installer log;
WARNING: The following NVIDIA GPUs are supported through NVIDIA legacy Linux graphics drivers and will be ignored by the NVIDIA 590.48.01 Linux graphics driver:
NVIDIA GeForce GTX 1070 requires 580.xx
This would suggest that I cant have that cuda126 without the 590 nvidia driver, and that the 590 driver wont install with my card? Ah, that warm, familiar feeling of going round in circles again..
will never use windows again.. ever.. really dont want to get into that rabbit hole here but for me, suffice to say, while these big corps are run by.. the sort of people they are run by, they wont ever get a penny out of me..
So its Linux, and if this ai stuff is not ready for us noobs to enjoy yet, then its back to gimp for me..
Okay. I get where you’re coming from. For me, an OS is just a tool, and Windows often saves me a lot of time and clicks, which makes life easier.
As expected, the problem is your card here. The driver does not support this old card anymore. Try NVIDIA 580.xx
So stick to the cu126, but with 580 drivers? Installer is for 590 only, will try and find another one..
No, you need to downgrade Cuda too. That would be cu 113 or 116. And being on Linux makes things even more complicated. This card is simply not longer supported.
Honestly, i would buy another card / pc instead. Comfy does not make fun with such ancient hardware. The timing is a bit bad at the moment though. The hardware economy is crazy at the moment.
Ah… so it was never going to work? Thank you..! think you may have saved me a lot of time flogging this dead horse and getting nowhere..
Managed to get jan ai and local ai working, seemingly without the problems I am getting with comfyui.. Thanks for help guys, time to move on..
1 Like
Finally tracked down the problems and found a solution.
I had been using Xubuntu which uses the xfce desktop. For some reason, the only nvidia drivers that would install were 570, 580, and 590. The 590 does not work at all (blank screen), 570 would not install no matter what I tried, which left 580 which did install but made the menus and non-full screen windows flicker, sometimes so badly I had to reboot. Whatever the errors were with this driver, they prevented the right cuda drivers installing properly which led to all the other problems. I am no expert in Linux but the above is my best guess.
The solution.. Discovered this by accident, but it works and I have zero problems now. I installed Ubuntu (gnome desktop), installed the nvidia drivers and this time I was offered the usual range of drivers (490-570) and the system automatically chose 535. This installed without any issue and so did cuda, docker and everything else that was producing errors before.
Once the system was configered, I installed and switched to Xcfe, rebooted, and there was my system, drivers working and all the ai software working too..
Hope this helps anyone else with a gtx 1070 having similar problems, I do frown a little more at my nvidia card now and have begun eyeballing an amd card for my next upgrade, nvidia drivers are still not open source and every time I have a problem and go looking on the net for a fix, I am always struck by the sheer numbers of people out there fighting with nvidia drivers..
1 Like
Cool that you finally found a solution. Wouldn’t have thought about that 
I would keep my hands off AMD, at least until the slowdown issues are resolved. For Comfy, you really want an Nvidia card.
I made that mistake myself. I bought a high end AMD card that wears the AI in the name a few weeks ago and I regret it. You can get it to work. I currently run Comfy in Docker on Ubuntu and natively on Windows. But I still have major problems, including dramatic slowdowns. Quite a few nodes does simply not work with an AMD card too, relying at the CUDA architecture.
It is not fun having to restart Docker or even the whole system after generating a single image or video. These slowdown bugs have been reported, but I do not have high hopes. As far as I know, these issues have been around for quite some time.