pc turns off while generating

[ExTroll1113]

when im generating my pc turns off

i have a rayzen 7 3700x
amd 590rx
16 gb ram ddr4
and my psu have 800 wats is a mars mp 800, never had issues with it.

MY COMMANDLINE_ARGS=--medvram --precision full --no-half --no-half-vae --disable-nan-check --opt-sub-quad-attention

[viebrix]

Sometimes I have the same problem since update to version 1.6.
I have AMD 580 RX and also a AMD Ryzen 7 3700X with 32 MB PSU is only 650W
Some in the forums think it happens when the memory is full and therefore restarting webui periodic to free vram
others think they have the problem because of overheating the gpu

I'm now testing with those startparameters (have read from them in another forum):
export COMMANDLINE_ARGS="--medvram --no-gradio-queue --opt-sub-quad-attention --sub-quad-q-chunk-size 256 --sub-quad-kv-chunk-size 256 --sub-quad-chunk-threshold 70 --no-half-vae --opt-split-attention --medvram-sdxl"
export PYTORCH_CUDA_ALLOC_CONF="garbage_collection_threshold:0.6,max_split_size_mb:128"
export SAFETENSORS_FAST_GPU=1

[verdooft]

When you use Linux, you could see the reason for the shutdown with:
journalctl -b -1

[viebrix]

So it seems the CommandlineArgs do not help further. Yesterday I had again a crash of my PC. I used ADetailer and went out of the room to eat something. Later i returned and the PC was blowing again with black screen - I dont know how long.

@verdooft thanks for your hint with journal, I did it with parameter -r
journalctl -b -1 -r
to get the most recent entries because the pc run very long time.

I have tried to anonymize it and I also replaced repeated entries with ":"

I can see there is a problem with a try to reset amdgpu which didn't work, but I can not understand the cause - why the reset was tried.
journal_gpu_crash_anonym.txt

I found also something about the temperature of hdd but this was already after the crash but before the reboot
I see also that cinnamon desktop crashed. Maybe linux run further and only all gpu threads were killed?

thanks a lot for help

Edit:
I think it's important, that I never had this problem with the prev. version of web-ui, there I had sometimes out of memory messages and the web-ui stoped, but never that the pc crashed.

[verdooft]

I had amdgpu issues too, but they are gone, since i use:

• recent kernel (Mainline or self compiled)
• recent mesa
• more recent package xserver-xorg-video-amdgpu
  Perhaps you are using a distribution like Arch Linux with recent software.
  Example for the not ready kernel 6.6:
• a lot of fixes for the AMDGPU driver
  https://9to5linux.com/linus-torvalds-announces-first-linux-kernel-6-6-release-candidate

But i use CPU only for Stable Diffusion.

[viebrix]

Thanks for your hints and infos!
Maybe I will try a newer mesa driver. Maybe I will also again try to get rocm5.6 instead of rocm 5.5 to run. All tries currently led to the segmentation fault error.
I used Easydiffusion with cpu. There a 768x768px picture needed 5-15minutes.
Using web-ui 1.5.2 and rocm lead to a lot faster image generation, and there are possibilities I did not have with easy diffusion.

After upgrade to web-ui 1.6 - I mistaken started without installing my compiled rocm-pytorch version and therefore it started in cpu mode. A simple picture needed extremly long and with sdxl and controllnet I had also a pc crash.
I never had this with version 1.5.2 from web-ui, therefore I think there must be something new, a new optimization or something with the new cache, which causes the possibility to crash something if memory is too small.

[viebrix]

After updating mesa drivers and updating motherboard bios and many other tries, I gave up and rolled back to web-ui 1.5.2 There SDXL worked without PC crashes, but controlnet for SDXL was not compatible, therefore I wasn't able to use it. Could only use contzrol net on 1.6 which crashed my complete PC everytime after 2-5 generations.

Last Weekend I tried again to upgrade to web-ui 1.7 and it also crashed. Therefore I tried to upgrade pytorch (and vision) to last version (upgrade means compile on my pc for my rocm 5.5 version and install in web-ui).
Since then I had no PC crash - only sometimes it stops with Error:
/webui.sh: Zeile 256: 206947 Getötet

But my PC survives and I can restart web-ui and start generation again :)

If someone has same problem with GFX803 means RX 580 etc. on linux mint or ubuntu
here is how I did it:
https://github.com/viebrix/pytorch-gfx803

[Talk2Giuseppe]

These issues appear to be common on low memory cards. Here are a few things I have done with varying degrees of success.

First, keep in mind that since we are already working with limited memory, the smallest of changes can have significant impact - both positively and negatively.

• Various models will play a part on these issues. Some models cause CUDA errors, some simply crash A1111. Sometimes changing the model will relieve some pain. I believe SD keeps two models in VRAM. Also, I think I saw some command you can do to remove one from memory which can free things up a bit when you really need to use a certain model. If that does exist, maybe someone can speak more to that process.
• Same concept on the resolution and/or the number of steps. When all seems to be going wrong, drop down to 20 steps and a 512 x 512 output to see if that resolves some issues.
• Command line args help. The only thing you give up is speed. It should be the first fix to apply.
• If you are able to monitor the stats on your card during a render, do it. Keep an eye on memory used and watts. Both will tell you if you've gone to far and you're experiencing hardware issues rather than software issues. If you're tapping out on the memory, lowering the resolution and/or steps helps. Also, turning off extensions (ie: not using ADetailer for example) and/or limiting LoRAs can help too.
• Finally, if you notice the watts climbing and A1111 goes sideways at about the same wattage every time, it's most likely the PSU. It doesn't matter at that point what size PSU you have in the machine, it could be at EOL and simply need to be replaced.
• One last one, and I am still trying to figure this one out myself - mixing SD LoRAs with XL models seems to cause issues SOMETIMES. Again, I haven't been able to consistently reproduce the problem. I only notice it when I start to look into the prompt to see if it might be the cause of a collapsed session. This typically happens when I copy prompts from websites to reproduce someone else's images.

I hope some of these items help. Let us know what worked for you so that we can all benefit from it. Thanks!