How to install A1111 on MacOS (step by step guide) #2
Replies: 6 comments 26 replies
-
|
I downloaded the 0.0.2 a1111-setup script today and tried to run it, but it was unsuccessful: So i tried following the "existing install" instructions above instead. @viking1304 - I wonder if it would be a good idea for a1111-setup.sh to include those commands if an existing install was detected? After doing these, the Web UI loaded up successfully :) |
Beta Was this translation helpful? Give feedback.
-
|
My M1 shows almost 100% usage of GPU during generation. 
|
Beta Was this translation helpful? Give feedback.
-
|
I just released version 0.0.3, which brings some tweaks and minor improvements. |
Beta Was this translation helpful? Give feedback.
-
|
Weird I wonder why my Radeon has low utilization. I'm just thinking... I'd like to remove the PyTorch install I have (the developer build) and go with the one you recommend. How can I uninstall PyTorch (installed using pip as instructed here)? |
Beta Was this translation helpful? Give feedback.
-
|
Actually after some Googling I think I figured it out |
Beta Was this translation helpful? Give feedback.
-
580X 8GB in my iMac (with exactly the same params as my M1, which does not work for you)
|
Beta Was this translation helpful? Give feedback.
-
This is the exact reason why I do not change command line args by default anymore. Git reset hard should restore the original file from A1111, where params are not set at all, and then it will change the command line only if you use the Since I have a feeling that your installation is still a bit messy, please do this:
|
Beta Was this translation helpful? Give feedback.
-
So after uninstalling my dev version of PyTorch, and following the instructions above to create a fresh install of A1111, I now have it running on PyTorch 2.1.2 :) 
I'm getting brown/grey boxes generated on the first run of a fresh install of A1111 though, somewhat as expected, since we already determined last week that the default arguments don't work for my setup (perhaps due to me being on Apple Intel.... what we first discussed on Discord anyway), so I updated them. For this initial run of SD UI, Terminal said:
so I edited my webui-user.sh file and updated line 13 to:
i.e. changing After doing this, now no models will load in the UI for me! :( |
Beta Was this translation helpful? Give feedback.
-
You can run it from anywhere except
I thought this sentence was clear enough:
You had PyTorch installed as a global package, which usually is not a problem, but since it was newer than the one I specified, the development version was used. Now you can choose which version you want to use inside venv that A1111 uses.
Those are the default command line parameters for Mac that the A1111 team added, which, in theory, should work for any standard Mac. They will be used unless you specify something different in
I noticed that, in some cases, models couldn't load and produce errors when I switch from SDXL to 1.5 or vice versa if I do not switch VAE back to Automatic first. In this case, I recommend switching VAE back to Automatic and restarting webui before trying to change the model. |
Beta Was this translation helpful? Give feedback.
-
The same thing just happened to me on iMac with 580x. I solved the problem by adding Try those command lines, please:
|
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the tip here. Unfortunately adding There seems to be an issue with the model loading component of the Web UI now..... no model is being loaded on startup of the Web UI, and I cannot switch the model in the UI any longer. I don't understand why it cannot find I don't think it's the subfolder that's the problem, since this issue was occurring even before I put my models in subfolders, plus it's mentioned it works here. Suspecting it was symlink issue, I deleted the symlink and copied the actual "v1-5-pruned-emaonly.safetensors" model from my other drive to the /models/Stable-diffusion/ folder. After that, I still get the "cannot find model" error: .... but when I try to generate an image, nothing happens. And when I try to switch model, nothing happens. :( Is there a way for "flush" previously loaded models and have A1111 rescan all models again like new? I don't particularly wanna run the a1111-setup.sh script again since it took me a while to set up all my symlinks! :) |
Beta Was this translation helpful? Give feedback.
-
|
No worries - I did a complete reinstall of A1111 and am back up and running! |
Beta Was this translation helpful? Give feedback.
-
|
Hi @viking1304, First off, thanks for your step-by-step guide on getting A1111 up and running on my MACOS. I've found it really useful, but I'm getting an error with Steps 5 and 6 of the common steps. I've followed all the Steps before that for the new install. For Step 5, I'm encountering this error: And for Step 6, I'm running into this issue: I must admit, I've never really used the terminal – probably like five times in my entire life. So, any guidance on what I might be doing wrong would be greatly appreciated. |
Beta Was this translation helpful? Give feedback.
-
|
You need to be inside the stable-diffusion-webui folder for that to work. You probably closed the terminal after you did some of the previous steps, so the terminal sent you to your home folder when you re-opened it. To get to the proper folder, you need a command from the start: But... if you do not feel "at home" with the terminal, you can use my script, which will do everything for you. Just run this and wait a little: |
Beta Was this translation helpful? Give feedback.
-
|
Thank you very much for the quick reply! It has worked this time and I was able to run it! |
Beta Was this translation helpful? Give feedback.
-
|
Thanks a lot for the helpful script. How do I proceed with an update?
Do I run the script again? Will the directories with the stable diffusion models and the generated images retained? |
Beta Was this translation helpful? Give feedback.
-
|
That sounds strange. What does Yes, you can rerun the script, and it should fix everything. Just be sure to use the latest version since something new and strange happens constantly. It will not touch your models, loras or extensions, but it will remove your venv folder, and a1111 will create a new one. The only thing that might need to be changed manually later would be command-line parameters in webui-user.sh, since it will be reset to default. If you want to use my params (that works best for me) instead of the default ones, run the script like this |
Beta Was this translation helpful? Give feedback.
-
|
Thank you for your support Viking!
|
Beta Was this translation helpful? Give feedback.
-
|
It looks like you were one of those who reverted A1111 to v1.7.0 by force. This should fix the detached state and give you the latest version of A1111:
I will update my script to handle detached head cases soon. Do not use torch 2.2.0 or newer (including nighly) if you are on macOS 14.4. Euler A and similar samplers do not work properly in that environment. Everything works as expected with 2.1.2. |
Beta Was this translation helpful? Give feedback.
-
|
awesome script, thanks for making it! installing Forge with it significantly reduced time for image generation. but only for SD 1.5 models, SDXL models loading time is approx minute, moving takes around same time and 512x512px 20 steps image generation takes around 40 min. is there anything possible to get SDXL model working faster or just stick to SD 1.5? |
Beta Was this translation helpful? Give feedback.
-
|
What is your exact Mac model? Also, post the log from terminal, so I can check if everything is OK. |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for reply! I later also discovered that configuration (webui-user.sh) coming with installation was working only with 512x512 images. Anything above, even 768x768 on SD1.5 was causing Mac to freeze. I looked in some threads/forums for info, changing configuration to what it is now in log have solved this issue. 512x512 20 steps images are generated in less than minute, 1024x1024 20 steps is taking around 3-5 minutes. System info: Times for SD/SDXL 512x512 20 step image generations (simple prompt: Photo, Red apple on wooden table, outdoors): Log from terminal (base) UserName@Mac GenAI % cd Forge ################################################################ ################################################################ ################################################################ ################################################################ ################################################################ To create a public link, set |
Beta Was this translation helpful? Give feedback.
-
|
Do not use Try this instead: |
Beta Was this translation helpful? Give feedback.
-
|
Unfortunately with this configuration SDXL model still takes more than 20 min to generate 512x512 image: To load target model SDXLClipModel |
Beta Was this translation helpful? Give feedback.
-
|
Your main problem is RAM. I think you should stick with 1.5, and try lightning since it can generate good images in fewer steps (like 4 or 8). Also, do not use 512x512 with SDXL (except with turbo). Use 768x768, 1024x1024, 1024x768 or 768x1024. For params, try this as suggested in your previous log: and also this |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for the script, installation went without issues using the 0.1.0 script from 3 weeks ago. I have used A1111 some on PC, and decided it was time to try on MacOS. Running on Sonoma 14.4.1 on M3 Max w/36GB RAM, I've found that several of the sampling methods produce only static. To test, I'm just using the base 1.5 pruned. These samplers produce just static: I've tried a few variations of the command line args based on the mention of noise results, but those apparently were just for gpus on intel. Any suggestions on how to proceed? Should I try a different version of PyTorch? Also, does xformers work on arm macs? I'm getting conflicting information as I thought xformers was nvidia only, but some places mention it running on arm too. EDIT: OK, I'm not sure what fixed it, but I thought maybe it would be best to use the recommended Python version, so I backed down from 3.12 to 3.10.9. Maybe it was coincidence, but launching webui.sh, it downloaded a new torch nightly version, and the broken samplers are working. I assume once 2.4 is stable, we can switch from nightly to stable? |
Beta Was this translation helpful? Give feedback.
-
March 17th 2024 version
UPDATE:
IMPORTANT NOTE - since the macOS 14.4 update, in combination with PyTorch 2.2.0 and newer, breaks some samplers like Euler A, DPM++ 2S A and some others, PyTorch 2.1.2 should be used.
If you want to install a1111 or fix your current a1111 install as fast as possible, just use my a1111-setup script that will do everything for you.
Simply run
bash a1111-setup.shorchmod 755 a1111-setup.sh && ./a1111-setup.sh. Wait a few minutes (depending on your computer and internet speed) until you see WebUI opened in your default browser.I have also added a few command line parameters:
Of course, you can still use this step-by-step guide if you want full control.
Step by step guide
This guide uses the best command line settings I could find after a lot of experimenting. It also solves all the errors I experienced with A1111. It is based on the original post from this thread. I am just posting everything together for easier access to proper and updated information.
It was tested and has been working without a problem on those Mac models:
DO NOT try to install the AMD version on iMacs with Radeon since Mac uses Metal even for Radeon Cards
New install:
If Homebrew is not installed, follow the instructions at https://brew.sh to install it. Keep the terminal window open and follow the instructions under "Next steps" to add Homebrew to your PATH
Open a new terminal window and run
brew install cmake protobuf rust python@3.10 git wgetType
cdto be sure that you are in your home folderClone the web UI repository by running
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webuiOptional step - place Stable Diffusion models/checkpoints you want to use into
stable-diffusion-webui/models/Stable-diffusion. If you skip this step, the default model will be downloaded.Optional step - A1111 v1.8.0 introduced some breaking changes, which might affect clean installations of A1111. If you just installed A1111 and plan to use Controlnet, please do those additional steps:
venv/bin/pip install protobuf==3.20.3venv/bin/pip install insightface==0.7.3Existing install:
Go to the stable-diffusion-webui folder using
cd stable-diffusion-webuiUse
git reset --hard origin/masterto remove local changes of AUTOMATIC1111 filesUse
git pullto update to the latest version of AUTOMATIC1111Run
venv/bin/pip cache purgeto clean pip’s wheel cacheRemove venv folder with rm -rf venv or using Finder.
Common steps:
Please check if you still have
venvfolder inside your stable diffusion folder. If you have it, please go back to the Existing install part and do everything from that part fistOpen webui-user.sh, find this line
and add the line below to use PyTorch 2.1.2
In the same file, find this line
and add the line below to use the recommended command line parameters for Mac
Your webui-user.sh should look like this:
Note: Do not edit
webui.shorwebui-macos-env.shSince some models might give you this error
TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.run this command to fix that error in advance
Run
./webui.sh. A new Python virtual environment will be created and activated using venv and all missing dependencies will be automatically downloaded and installed. Web UI will be launched automatically at the endAdditional notes:
To relaunch the web UI process later, run
./webui.shagain. To update, rungit pullbefore running./webui.shIn some cases, default command line parameters provided by the A1111 team are not the best choice, and in some rare cases, they do not work.
You can try and see which of the command line parameters below works the best for you.
Args for Macs implemented by the A1111 team:
My recommendation for most Macs:
For special cases when
--no-half-vaeproduce only noise:Those combinations also worked without error on my Macs:
Beta Was this translation helpful? Give feedback.
All reactions