How to enable mps on Mac?

[shark-speare]

I’m using a MacBook with the Apple M3 chip.

My launch parameters are:

--skip-torch-cuda-test --upcast-sampling --opt-sub-quad-attention --medvram

I’ve also installed the nightly version of PyTorch.

However, generating a single 1024×1024 image with 20 steps takes around 20 minutes, which seems abnormally slow.
I did some research and found that I should be able to enable MPS (Metal Performance Shaders). When MPS is working, the log should show Using torch device: mps.

But no matter what I try, I can’t get it to work — the log never shows that MPS is being used.
Has anyone encountered this on an M3 Mac? Any idea how to fix it?

[reekes]

I have decent performance on my M2 Ultra. Not as good as a cheap PC with a video card, but it's acceptable. Here's what I have in my
webui-user.sh file

export COMMANDLINE_ARGS="--skip-torch-cuda-test --opt-sub-quad-attention --no-half-vae --no-half --theme=dark"

It also took me many weeks to work out how to optimize my environment. I use python 3.10.17 and torch 2.7.1.

Also in webui-user.sh I have this

export TORCH_COMMAND="pip install torch torchvision"

If you make a change to your python environment, make sure you delete the /venv directory and then run ./webui.sh to pull the packages. Double check at the bottom of the web page you see the correct versions. I use python 3.10.17 and torch 2.7.1.

I tried versions of torch after 2.7.1 but they all produced garbage. You also need Python 3.8 or later.

I can get 3.6 to 4.4 it/s depending on which model I use.