CUDA out of memory (SOLVED)

[Elonel5]

I'm encountering a problem with installing: (followed all the retard instructions like a good retard)

RuntimeError: CUDA out of memory. Tried to allocate 20.00 MiB (GPU 0; 4.00 GiB total capacity; 3.42 GiB already allocated; 0 bytes free; 3.48 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

and installation loops
Anything I can do?

[EliEron]

The message is telling you that you ran out of VRAM. To run Standard-Diffusion on low VRAM cards you currently have to enable one of the optimized modes. The standard optimized mode lowers VRAM a lot but also massively affects speed. The optimize-turbo mode uses a bit more VRAM but runs with little speed penalty.

To enable the optimized mode open the "relauncher.py" file and change:

optimized = False

To:

optimized = True

To enable optimized-turbo do the same thing but for the "optimized_turbo" option. Though only one of them should be enabled (set to True) at a time.

With only 4GB of VRAM (based on the error message) you will likely have to use the standard optimized mode for now, but there is a PR open right now that will lower VRAM requirements a bit further once merged without a speed penalty. And another that should lower it even further but with a bit of a speed penalty. So you might be able to switch to optimized-turbo in the near future

[EliEron]

Ah I got you, that makes more sense. I didn't quite get that from reading your comment the first time, though rereading it I can see where you were coming from. I might give that a try later, as I do usually place the main subject at the front in the prompt.

[Elonel5]

It worked! thanks!

[denizbinay]

In fact, I experiment a lot with patterns and the results are much better in 1024x1024 than in 512x512. Maybe because there is room for repetition.

[codefaux]

I'm receiving a CUDA out of memory error as well, but my situation is a bit different.

A few years back, my partner and I built two identical high-spec gaming machines. The case is different, but the GPU (1080 TI), motherboard, CPU (i7-8600k), RAM (32gb), PSU, SSDs, (you get it) are all identical. My machine runs SD in Linux, his is untested in Linux due to Windows preference. Both machines also run Windows 10 64-bit, same update channel.

My machine works flawlessly with Stable Diffusion in all modes. His crashes during init with a Cuda out-of-memory, before the port is even opened. We've cloned my install to his system, to make sure it wasn't a corrupt download. Fresh install, same steps on both systems. Drivers should match. Ideas?

[mchaker]

@codefaux yes, possibly:

Windows could be using some (or more) GPU capacity to render the desktop environment (resource usage would increase when monitor size increases), reducing the available resources for stable-diffusion.

[mchaker]

Have you tried the Doggettx optimization?

Item 1 in this list: https://gist.github.com/mchaker/e0cf180db6849fa03696293b3fc8f485

[Elonel5]

Thank you for your patience, I don't come from programming and I'm also new to GitHub

I entered CompVis/stable-diffusion@main...Doggettx:stable-diffusion:main

How do you I download the new files?
I see the numbers on the side, that look like the index number, but can't find d8ad218c

Thank you so much

[Elonel5]

I'm actually still lost, can't find d8ad218c, How do I download the new files?

[mchaker]

Apologies - so, to download the new files:

1. Go to the link from your post (Doggettx optimizations)
2. Scroll down until you see the file header for attention.py:
   
3. on the right side of that file header, you should see a 3-dots menu
4. click the 3-dots menu and select "view file"
   
5. On the right side of the page, you should see a "Raw" button. Click that
   
6. On this page (it should just be a page of code), press Ctrl-S (or Command-S on Mac) to save the page as a file. Save this attention.py file under your stable diffusion folder, inside the following path: (stable-diffusion folder)/ldm/modules/ .
7. Repeat steps 2 through 6 for the next file: model.py, placing it in the (stable-diffusion folder)/ldm/modules/diffusionmodules/ folder.
   

[Elonel5]

Thank you so much KING! Did this, and rendered a 512 x 704 !!!
I appreciate the detailed instructions and could have never figured it out myself!

[mchaker]

Haha thanks and awesome!!! Glad to hear it :)