Why am I getting RuntimeError: Invalid buffer size: 3.57 GB with 16GB Vram?

[TijuanaKez]

So wierd problem running A1111 Stable Diffusion on macOS Ventura. I can generate up to 640 x 640 images fine, but above that I get errors.

For example 704 x 704 generates the RuntimeError: Invalid buffer size: 3.57 GB

If I try the same in InvokeAI, I create 1024x1024 images on this GPU no problem, probably higher haven't tried.

I'm using a RT 6950 XT with 16GB VRam, (64GB system ram) and SD is running via MPS.

I can see the Radeon Card memory and processor being used in iStat Menus(and its fans spin up to loud mode) when I'm generating an image (and my CPU cores stay around %2) so I know its running on the GPU via MPS, not CPU.

Choice of model, whether it's based on SD1.5 or SD2.0, or choice of sampler makes no difference.

Is there some other command line args I should be using?

(happens with and without --no-half)

Python 3.10.11 (main, Apr  7 2023, 07:31:31) [Clang 14.0.0 (clang-1400.0.29.202)]
Version: v1.3.1
Commit hash: b6af0a3809ea869fb180633f9affcae4b199ffcf
Installing requirements
Launching Web UI with arguments: --skip-torch-cuda-test --upcast-sampling --opt-split-attention
No module 'xformers'. Proceeding without it.
Warning: caught exception 'Torch not compiled with CUDA enabled', memory monitor disabled

[Bs-2195]

SD1.5和SD2.0是什么意思 哥

[xiaodeng111]

Have you solved your problem?

[hqnicolas]

@TijuanaKez I have an REPO that solves AMD ROCm RDNA2 & 3 problems with docker containers on linux
https://github.com/hqnicolas/StableDiffusionROCm
you cold try to use this containers
it was stable at 1.9.3 (latest)
if you like this automation REPO please let a Star on it ⭐