[bug]: MPS support for AMD GPUs under macOS - grey blank image result output

[neborion]

Is there an existing issue for this?

• I have searched the existing issues

OS

macOS

GPU

amd

VRAM

8GB

What version did you experience this issue on?

3.4.0post2

What happened?

When I use the CPU, the generation works. It's extremely slow. 8s/it
When I use MPS or leave it on "auto", it's fast 1.05s/it. On the other hand, it only produces an empty, brown, uniform image, a grey hazy fog. That's why I put the "bug" tag.
Browsing the web, I'm having great difficulty determining whether MPS works with AMD GPUs on Macs. So my question is:

is there a way to get AMD GPU acceleration working on macOS?

What do you think of the fact that InvokeAI tries to use MPS when set to auto? It had given me hope that it would work on AMD GPUs.
The machine is an i7-6700, 32GB RAM, AMD 5700XT 8go VRAM (RDNA1), macOS 13.3 Ventura, pyenv & python 3.10.11
I solved all the warnings I could in terminal. In case I'm misdiagnosing and my problem isn't MPS incompatibility on AMD GPUs, here are the warnings I have left in console:

/InvokeAI/.venv/lib/python3.10/site-packages/torchvision/transforms/functional_tensor.py:5: UserWarning: The torchvision.transforms.functional_tensor module is deprecated in 0.15 and will be **removed in 0.17**. Please don't rely on it. You probably just need to use APIs in torchvision.transforms.functional or in torchvision.transforms.v2.functional.

objc[21114]: Class CaptureDelegate is implemented in both /InvokeAI/.venv/lib/python3.10/site-packages/cv2/cv2.abi3.so (0x14e74a648) and /usr/local/Cellar/opencv/4.8.1_4/lib/libopencv_videoio.4.8.1.dylib (0x160988898). One of the two will be used. Which one is undefined.

Thanks for your time

Grey outputs :

Screenshots

No response

Additional context

No response

Contact Details

No response

[Millu]

Is there a way to get AMD GPU acceleration working on macOS? / What do you think of the fact that InvokeAI tries to use MPS when set to auto? It had given me hope that it would work on AMD GPUs.

I believe this is a result of Invoke doing a naive check for Darwin and then assuming MPS. Unfortunately support for Intel Macs is rather limited

In case I'm misdiagnosing and my problem isn't MPS incompatibility on AMD GPUs, here are the warnings I have left in console:

These don't appear to be relevant to the issue you're experiencing 😞

[eighthkeepa]

I actually had it running pretty well with Invoke 3.0.2post1 on an iMac with Monterey a while ago.
Now I have Invoke 3.6.0 on Ventura, same mac, and it gives me the same brown images as on @neborion screenshots with mps enabled.

[neborion]

I actually had it running pretty well with Invoke 3.0.2post1 on an iMac with Monterey a while ago. Now I have Invoke 3.6.0 on Ventura, same mac, and it gives me the same brown images as on @neborion screenshots with mps enabled.

@eighthkeepa, In my case, there wasn't really a bug. It's just that acceleration via AMD RDNA1 GPU is not available (at the time of this post). Sad. So you have to switch to CPU usage. When starting invokeAI, in the terminal, select option n°5 "InvokeAI startup option". Then, on the "Generation device" line, select "CPU". Restart invokeAI.
Hope this helps. Best regards.