InvokeAI [bug]: Cuda out of memory during loading inpainting model

[bug]: Cuda out of memory during loading inpainting model

Open BartaG512 opened this issue 1 year ago • 7 comments

Is there an existing issue for this?

[X] I have searched the existing issues

OS

Windows

GPU

I dont know

VRAM

No response

What happened?

I choose inpainting v5 model and the loading failed

Screenshots

Additional context

No response

Contact Details

[email protected]

Dec 04 '22 08:12 BartaG512

I got the exact same error and in my bug report I got this response: "In the latest version we activate the NSFW checker by default and this eats up an additional ~0.5GB of memory. If you've got the NSFW checker on, you can try turning it off. Find the .invokeai file located in your \User directory and change the line that reads --nsfw_checker to --no-nsfw_checker." But I still don't know how to find the file in the directory they were talking about. Do you know how to fix it?

Dec 04 '22 11:12 FilipTrichkov

I turned it off You can find it in "C:/Users/username/.invokaai" file But still gettting this error.

Dec 04 '22 15:12 BartaG512

it would be helpful if you could post what GPU your system has. Several ways to do that:

Open the “Run” dialog box by pressing and holding the Windows key, then press the R key (“Run”). Type devmgmt.msc Click OK. Click on Display adapters
Open Task manager: right click the windows task bar and click on "Task Manager" go to the second tab (performance) and look for the "GPU 0" entry. Below GPU 0 it should say what GPU you have,

With that information we can tell you if you have a fixable problem or just a too weak of a system to run InvokeAI.

Dec 04 '22 15:12 FalxGod

GPU 0 Intel(R) UHD Graphocs 630 GPU 1 NVIDIA GeForce GTX 1050 Ti

Maybe should I specify to use GPU 1? How can I do that?

Dec 04 '22 16:12 BartaG512

I've got several Out of memory errors as well using inpainting, (6gb here) as a strange workaround I found that switching the model to a different one (like the v1.5) and back again to inpainting solve the out of memory for many runs... You would think that more model is more ram but maybe the cache kicks in and some memory is freed/optimized in the process. (and before you ask nsfw is on and should be on by default I don't like censorship by default!)

Dec 04 '22 22:12 diramazioni

BartaG512 the command to do that is : set CUDA_VISIBLE_DEVICES=1 I was using that in automatic1111 however, in invoke, you are supposed to put it in the c:/users/username/ invokeai I tried that and only get errors, it's not working for me. if someone can help please ? I tried a more simple command of COMMANDLINE_ARGS=--opt-split-attention --medvram and also could not get this to work.