"Not enough memory resources are available to complete this operation"

[ChaseFlorom]

Hey, I'm running a Radeon RX 5700 with 16GB and I'm getting a not enough memory resources are available to complete this operation. Any ideas? I've had the CMD version of SD running and had it work.

Thanks!

[azuritecoin]

Are you running with a too many batch size or too large image? I find that on my RX 6900 XT 16GB i can run up to a batch size of 3 512x512 images or a single image size of 704x704. 768x512 is ok but slow. Try closing the UI prrogram and restarting it.

[ChaseFlorom]

It was totally a me problem for this one. I had messed with my comp the night before and one of my sticks of ram was partially inserted wrong.

[Dolyfin]

Having similar problem with vram. 12GB 5700XT, Batch size 1 and Iteration 1 seems to already be at the limit with 512x+ resolution. Also, it appears that I have to restart onnxUI.py every few runs because the VRAM would fill to max and never be cleared when idle. Eventually VRAM will be full enough to not allow to start unless exiting onnxUI.py which clears VRAM.

[azuritecoin]

Try a smaller image size of 384x384, or share more of your system RAM with the GPU. From my own experience, 16GB of total memory (VRAM + shared system memeory) should be enough for batch size of one 512x512 image.

[amida168]

I had this problem too. I have a laptop with hybrid GPUs. I changed the Graphics Settings to solve this problem.

http://www.kombitz.com/2022/11/01/how-to-fix-stable-diffusion-not-enough-memory-resources-are-available-problem-on-windows-11/

[mikwee]

Same problem with my AMD Radeon RX 6600 and 8GB of RAM, in both 384x384 and 512x512. Batch size was 1 in both attempts.

[mikwee]

Update: 384x384 works with the CLI, 512x512 doesn't.

[azuritecoin]

Yeah, 8GB is the bare minimum that's needed to run stable diffusion. 512x512 needs more VRAM or shared system RAM.

[mikwee]

Yeah, honestly makes sense. 384x384 is impressive as is

[Tjololo12]

I run into this issue when trying to do two generations one after the other. It appears, looking at memory utilization, that the UI isn't releasing the memory from the previous batch. Is there a setting or something I need to enable to force a garbage collect after generation finishes? 6700XT

[azuritecoin]

The latest version of onnxUI.py has forced garbage collection after changing the pipeline. If you're willing to edit the script you could try adding gc.collect() at the begining and the end of generate_click function.

[WalterShomer]

which lines would that be? not programming savvy and tried a few combinations that didn't seem to work

[azuritecoin]

In onnxUI.py there's a function called run_diffusers around line 23. Scrolling down a bit, at line 42 you will see this global pipe.

Replace that line with the following. Make sure there's four spaces at the front of each line:

    global pipe
    pipe = None
    gc.collect()

This will unload the pipeline and free up the memory. The downside is that everytime you run this it will have to reload the pipeline which takes some time.

[azuritecoin]

I've updated the UI to enable de-allocating the model and releasing the memory after each generation. The downside is that you need to reload the model at the start every time. Run the UI with the flag python onnxUI.py --release-memory

[IslandRhythms]

I've had a similar issue and I believe the problem is the addition of the scheduler. If you could add an option to the GUI where no scheduler is selected, I believe this may solve the problem.

[azuritecoin]

I've ran a test with and without the user specified scheduler and it doesn't affect the amount of memory used. I end up using 11.3/16 GPU memory after generating one 512x512 image for both scenarios.

[IslandRhythms]

I don't know what to tell you. I have a AMD RX 6800 XT and when I run the non GUI script it works fine. When I transition the GUI it tells me it doesn't have enough memory to run the operation. I've commented out the original line and put in a line where no scheduler is included as an arg and it works fine

[azuritecoin]

When did you setup stable diffusion on your PC? The current version of the UI doesn't need the scheduler flag when you call onnxUI.py

[IslandRhythms]

https://github.com/azuritecoin/OnnxDiffusersUI/blob/main/onnxUI.py#L217
What is the default scheduler that from_pretrained uses? I'll try again using that but when I changed that line to
pipe = OnnxStableDiffusionPipeline.from_pretrained(model_path, provider=provider) I don't have an issue

[azuritecoin]

It depends on which model you load. eg. SD v1.5 will use PNDM by default. You can look in the \scheduler folder to see which scheduler your model uses.