HiRes Fix Seems to be getting stuck

[Helldiver1080]

Hey everyone, I am not sure if this is a bug or not as I am very new to Automatic1111. I am running a 3070ti, which has 8gb of VRAM. I am trying to use text2img and use the hires fix set at 2 for 2 images in a batch. The system gets to 50% and just hangs. No errors in the console, nothing printed in the UI, it just seems to sit there at 50%. Has anyone else experienced the HiRes Fix getting stuck or locking up like this?

[KolenoNakashi]

I think at 50% there is starting the upscale, try different upscaling method.

[BetaDoggo]

I'm not sure if this will help but I wouldn't recommend using a batch size above 1 with hires fix because higher batch size doesn't have a major effect on speed and it increases vram usage which will limit the size you can upscale to.

[Squeezitgirdle]

4090 here. This issue recently started happening for me too.

Used to be able to run hires fix for 50 batch for 3k images, now I can't do a single image above 2500 x 1400

[Squeezitgirdle]

50 batch?????????

batch size or batch count?
batch count, I would sometimes run 50 when I go to bed and hope I got lucky in the morning for some of the more difficult prompts I've run.

That said, for my use I don't usually need images larger than say 2560 x 1440 or so. But the reason I'm generating larger is so that I can make it more obvious how bad this issue is on a 4090.

I either fail or it takes a very long time (roughly 5 minutes compared to the 30 seconds it used to take) to generate an image around 2560 x 1440

4k is impossible now, though I used to be able to do it.

Basically it gets stuck with 15 seconds remaining right when hires steps starts, if it doesn't cuda error out, it will then continue, with x-formers it's not that bad though I didn't used to use it. Once it gets to 100% completion, it hangs again for a while and it will either fail there or it will take another minute before it finally gives me my image.

Image previews is turned off.

[w-e-w]

@Squeezitgirdle what cross attention optimization are you using?
so in 1.3.0
AUTO change the default cross attention optimization to SDP, before it was using Doggettx
as it's supposed to have better performance (average depending on the system)
the downside is it's less memory efficient so more likely to OOM

this seems to have caused lots of existing users who are using the default settings to OOM
so we reverted the default back to Doggettx in 1.3.1

you can try out different optimizations
see it helps with the memory situation

[Squeezitgirdle]

@Squeezitgirdle what cross attention optimization are you using? so in 1.3.0 AUTO change the default cross attention optimization to SDP, before it was using Doggettx as it's supposed to have better performance (average depending on the system) the downside is it's less memory efficient so more likely to OOM

this seems to have caused lots of existing users who are using the default settings to OOM so we reverted the default back to Doggettx in 1.3.1

you can try out different optimizations see it helps with the memory situation

I replied to you in another spot, but just in case someone with the same issue comes to this post, I am currently on xformers as its the one that works best. doggettx is significantly slower and fails more frequently. sdp - scaled dot product works almost as well as xformers but not quite.

[w-e-w]

the plot thickens............

[Squeezitgirdle]

the plot thickens............

?

[neavo]

Starting with some recent version (maybe v1.3.0?), I have the same problem. When doing image generation at larger resolutions -- on my system it's HiresFix or I2I at 1440x1920 -- the system regularly gets stuck.
STUCK means that the progress bar does not move at all and there is no output or errors in the console. If you are willing to wait five minutes or more, the system can still generate the target image normally, when the system monitor shows that the GPU power consumption is only about 30% of the normal full-load operation, which is obviously abnormal.

----- UPDATE -----

Probably the reason is not the WebUI update, but the update of nvdia driver. After I uninstalled the latest 535.98 driver using DDU in safe mode and then reinstalled the 532.03 driver, the problem seems to be solved.

[Squeezitgirdle]

interesting, I can give that a try but it'll be a while before I can

[YPL73]

this didnt work for me btw, this issue still occurs, i couldnt launch WebUI when i downloaded the 532.03 drive, but i still managed to downgrade my driver to 535.98

[YPL73]

interesting, I can give that a try but it'll be a while before I can

did this work for you?

[neavo]

see this

#11062

[DmytroSokhach]

I see people downgrade driver
But have the same problem with Driver Version: 545.84 10/17/2023 on RTX 3090.
Any other possible solutions?