Open Mozoloa opened 1 year ago
Since changing my settings to display every other step using TAESD I haven't had a freeze!
9-16-2023 This absolutely worked for me. Been having issues since reinstalling and i guess changing those settings, a few day ago. Single image generations randomly freezing, in the ui, the console, or both. Deforum animations randomly freezing. Changed this yesterday to 5 frame preview and Approx NN and havent had an issue since. This was the fix for me~
Just so we're clear, this is not a fix for the full preview, it's just using another preview engine, that's not full and it shows, and we already knew the other ones worked. It's not a solution
This isn't it either, at least for me. Watching the traffic go through when the interval is set to 1ms and 9999ms yields the same result - the last packet stalls due to python.exe hanging upon image completion. If packets throughout the image build stalled in the same fashion with the 1ms interval, that would support your theory.
This has become an issue for me recently. It happens every gen now.
Renaming venv and changing lines 318 and 319 in \modules\launch_utils.py
from:
torch_index_url = os.environ.get('TORCH_INDEX_URL', "https://download.pytorch.org/whl/cu121") torch_command = os.environ.get('TORCH_COMMAND', f"pip install torch==2.1.2 torchvision==0.16.2 --extra-index-url {torch_index_url}")
to: torch_index_url = os.environ.get('TORCH_INDEX_URL', "https://download.pytorch.org/whl/cu117") torch_command = os.environ.get('TORCH_COMMAND', f"pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 --extra-index-url {torch_index_url}") and if you use xformers, line 343 from: xformers_package = os.environ.get('XFORMERS_PACKAGE', 'xformers==0.0.23.post1') to xformers_package = os.environ.get('XFORMERS_PACKAGE', 'xformers==0.0.16rc425')
fixed the hanging at the cost of a pretty big performance hit. Would like a better workaround since I go from around 2.7it/s to 1.4it/s on my card.
Update: I downloaded the fp16 SDXL VAE here https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl.vae.safetensors
and removed --no-half-vae
from my launch arguments and added --medvram
. Problem is seemingly gone now so the above version change fix can be ignored if this works for you also.
Is there an existing issue for this?
What happened?
Since the update 1.1, very often when I do batches of images, one of them will hang at one of the latest steps and never complete.
Clicking interrupt does nothing, so does skip and reloading the UI doesn't help, the whole UI is stuck and it seems that no other functionality works. The console shows the total progress this way (I'm generating 100 batches of one 512x512 images ) :
I can't do anything but start the whole thing
Steps to reproduce the problem
What should have happened?
The generation should have continued like it did before
Commit where the problem happens
c3eced22fc7b9da4fbb2f55f2d53a7e5e511cfbd
What platforms do you use to access the UI ?
Windows 11, RTX3090
What browsers do you use to access the UI ?
Brave
Command Line Arguments
List of extensions
ControlNet v1.1.134 Image browser
Console logs
Additional information
I remember that at some point it hanged but got unstuck somehow and I got an error which I don't remember but it did say to use --no-half-vae, I haven't tested that and never needed that before on torch 1.13.1 for tens of thousands of gens. I'm exclusively using the new 840000 mse VAE