InvokeAI/README-Mac-MPS.md

macOS Instructions

Requirements

• macOS 12.3 Monterey or later
• Python
• Patience
• Apple Silicon*

*I haven't tested any of this on Intel Macs but I have read that one person got it to work, so Apple Silicon might not be requried.

Things have moved really fast and so these instructions change often and are often out-of-date. One of the problems is that there are so many different ways to run this.

We are trying to build a testing setup so that when we make changes it doesn't always break.

How to (this hasn't been 100% tested yet):

First get the weights checkpoint download started - it's big:

Sign up at https://huggingface.co Accept the terms and click Access Repository: https://huggingface.co/CompVis/stable-diffusion-v-1-4-original Download sd-v1-4.ckpt (4.27 GB) and note where you have saved it (probably the Downloads folder) While that is downloading, open Terminal and run the following commands one at a time.

# install brew (and Xcode command line tools):
/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/HEAD/install.sh)"

# install python 3, git, cmake, protobuf:
brew install cmake protobuf rust

# install miniconda (M1 arm64 version):
curl https://repo.anaconda.com/miniconda/Miniconda3-latest-MacOSX-arm64.sh -o Miniconda3-latest-MacOSX-arm64.sh
/bin/bash Miniconda3-latest-MacOSX-arm64.sh

# clone the repo
git clone https://github.com/lstein/stable-diffusion.git
cd stable-diffusion

#
# wait until the checkpoint file has downloaded, then proceed
#

# create symlink to checkpoint
mkdir -p models/ldm/stable-diffusion-v1/
PATH_TO_CKPT="$HOME/Downloads"  # or wherever you saved sd-v1-4.ckpt
ln -s "$PATH_TO_CKPT/sd-v1-4.ckpt" models/ldm/stable-diffusion-v1/model.ckpt

# install packages
PIP_EXISTS_ACTION=w CONDA_SUBDIR=osx-arm64 conda env create -f environment-mac.yaml
conda activate ldm

# only need to do this once
python scripts/preload_models.py

# run SD!
python scripts/dream.py --full_precision  # half-precision requires autocast and won't work

The original scripts should work as well.

python scripts/orig_scripts/txt2img.py --prompt "a photograph of an astronaut riding a horse" --plms

Note, export PIP_EXISTS_ACTION=w is a precaution to fix conda env create -f environment-mac.yaml never finishing in some situations. So it isn't required but wont hurt.

After you follow all the instructions and run dream.py you might get several errors. Here's the errors I've seen and found solutions for.

Is it slow?

Be sure to specify 1 sample and 1 iteration.

python ./scripts/orig_scripts/txt2img.py --prompt "ocean" --ddim_steps 5 --n_samples 1 --n_iter 1

Doesn't work anymore?

PyTorch nightly includes support for MPS. Because of this, this setup is inherently unstable. One morning I woke up and it no longer worked no matter what I did until I switched to miniforge. However, I have another Mac that works just fine with Anaconda. If you can't get it to work, please search a little first because many of the errors will get posted and solved. If you can't find a solution please create an issue.

One debugging step is to update to the latest version of PyTorch nightly.

conda install pytorch torchvision torchaudio -c pytorch-nightly

If conda env create -f environment-mac.yaml takes forever run this.

git clean -f

And run this.

conda clean --yes --all

Or you could reset Anaconda.

conda update --force-reinstall -y -n base -c defaults conda

"No module named cv2", torch, 'ldm', 'transformers', 'taming', etc.

There are several causes of these errors.

First, did you remember to conda activate ldm? If your terminal prompt begins with "(ldm)" then you activated it. If it begins with "(base)" or something else you haven't.

Second, you might've run ./scripts/preload_models.py or ./scripts/dream.py instead of python ./scripts/preload_models.py or python ./scripts/dream.py. The cause of this error is long so it's below.

Third, if it says you're missing taming you need to rebuild your virtual environment.

conda env remove -n ldm
conda env create -f environment-mac.yaml

Fourth, If you have activated the ldm virtual environment and tried rebuilding it, maybe the problem could be that I have something installed that you don't and you'll just need to manually install it. Make sure you activate the virtual environment so it installs there instead of globally.

conda activate ldm
pip install *name*

You might also need to install Rust (I mention this again below).

How many snakes are living in your computer?

Here's the reason why you have to specify which python to use. There are several versions of python on macOS and the computer is picking the wrong one. More specifically, preload_models.py and dream.py says to find the first python3 in the path environment variable. You can see which one it is picking with which python3. These are the mostly likely paths you'll see.

% which python3
/usr/bin/python3

The above path is part of the OS. However, that path is a stub that asks you if you want to install Xcode. If you have Xcode installed already, /usr/bin/python3 will execute /Library/Developer/CommandLineTools/usr/bin/python3 or /Applications/Xcode.app/Contents/Developer/usr/bin/python3 (depending on which Xcode you've selected with xcode-select).

% which python3
/opt/homebrew/bin/python3

If you installed python3 with Homebrew and you've modified your path to search for Homebrew binaries before system ones, you'll see the above path.

% which python
/opt/anaconda3/bin/python

If you drop the "3" you get an entirely different python. Note: starting in macOS 12.3, /usr/bin/python no longer exists (it was python 2 anyway).

If you have Anaconda installed, this is what you'll see. There is a /opt/anaconda3/bin/python3 also.

(ldm) % which python
/Users/name/miniforge3/envs/ldm/bin/python

This is what you'll see if you have miniforge and you've correctly activated the ldm environment. This is the goal.

It's all a mess and you should know how to modify the path environment variable if you want to fix it. Here's a brief hint of all the ways you can modify it (don't really have the time to explain it all here).

• ~/.zshrc
• ~/.bash_profile
• ~/.bashrc
• /etc/paths.d
• /etc/path

Which one you use will depend on what you have installed except putting a file in /etc/paths.d is what I prefer to do.

Debugging?

Tired of waiting for your renders to finish before you can see if it works? Reduce the steps! The image quality will be horrible but at least you'll get quick feedback.

python ./scripts/txt2img.py --prompt "ocean" --ddim_steps 5 --n_samples 1 --n_iter 1

OSError: Can't load tokenizer for 'openai/clip-vit-large-patch14'...

python scripts/preload_models.py

"The operator [name] is not current implemented for the MPS device." (sic)

Example error.

...
NotImplementedError: The operator 'aten::_index_put_impl_' is not current implemented for the MPS device. If you want this op to be added in priority during the prototype phase of this feature, please comment on [https://github.com/pytorch/pytorch/issues/77764](https://github.com/pytorch/pytorch/issues/77764). As a temporary fix, you can set the environment variable `PYTORCH_ENABLE_MPS_FALLBACK=1` to use the CPU as a fallback for this op. WARNING: this will be slower than running natively on MPS.

The lstein branch includes this fix in environment-mac.yaml.

"Could not build wheels for tokenizers"

I have not seen this error because I had Rust installed on my computer before I started playing with Stable Diffusion. The fix is to install Rust.

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh

How come --seed doesn't work?

First this:

Completely reproducible results are not guaranteed across PyTorch releases, individual commits, or different platforms. Furthermore, results may not be reproducible between CPU and GPU executions, even when using identical seeds.

PyTorch docs

Second, we might have a fix that at least gets a consistent seed sort of. We're still working on it.

libiomp5.dylib error?

OMP: Error #15: Initializing libiomp5.dylib, but found libomp.dylib already initialized.

You are likely using an Intel package by mistake. Be sure to run conda with the environment variable CONDA_SUBDIR=osx-arm64, like so:

CONDA_SUBDIR=osx-arm64 conda install ...

This error happens with Anaconda on Macs when the Intel-only mkl is pulled in by a dependency. nomkl is a metapackage designed to prevent this, by making it impossible to install mkl, but if your environment is already broken it may not work.

Do not use os.environ['KMP_DUPLICATE_LIB_OK']='True' or equivalents as this masks the underlying issue of using Intel packages.

Not enough memory.

This seems to be a common problem and is probably the underlying problem for a lot of symptoms (listed below). The fix is to lower your image size or to add model.half() right after the model is loaded. I should probably test it out. I've read that the reason this fixes problems is because it converts the model from 32-bit to 16-bit and that leaves more RAM for other things. I have no idea how that would affect the quality of the images though.

See this issue.

"Error: product of dimension sizes > 2**31'"

This error happens with img2img, which I haven't played with too much yet. But I know it's because your image is too big or the resolution isn't a multiple of 32x32. Because the stable-diffusion model was trained on images that were 512 x 512, it's always best to use that output size (which is the default). However, if you're using that size and you get the above error, try 256 x 256 or 512 x 256 or something as the source image.

BTW, 2**31-1 = 2,147,483,647, which is also 32-bit signed LONG_MAX in C.

I just got Rickrolled! Do I have a virus?

You don't have a virus. It's part of the project. Here's Rick and here's the code that swaps him in. It's a NSFW filter, which IMO, doesn't work very good (and we call this "computer vision", sheesh).

Actually, this could be happening because there's not enough RAM. You could try the model.half() suggestion or specify smaller output images.

My images come out black

We might have this fixed, we are still testing.

There's a similar issue on CUDA GPU's where the images come out green. Maybe it's the same issue? Someone in that issue says to use "--precision full", but this fork actually disables that flag. I don't know why, someone else provided that code and I don't know what it does. Maybe the model.half() suggestion above would fix this issue too. I should probably test it.

"view size is not compatible with input tensor's size and stride"

  File "/opt/anaconda3/envs/ldm/lib/python3.10/site-packages/torch/nn/functional.py", line 2511, in layer_norm
    return torch.layer_norm(input, normalized_shape, weight, bias, eps, torch.backends.cudnn.enabled)
RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...) instead.

Update to the latest version of lstein/stable-diffusion. We were patching pytorch but we found a file in stable-diffusion that we could change instead. This is a 32-bit vs 16-bit problem.

The processor must support the Intel bla bla bla

What? Intel? On an Apple Silicon?

Intel MKL FATAL ERROR: This system does not meet the minimum requirements for use of the Intel(R) Math Kernel Library.
The processor must support the Intel(R) Supplemental Streaming SIMD Extensions 3 (Intel(R) SSSE3) instructions.██████████████| 50/50 [02:25<00:00,  2.53s/it]
The processor must support the Intel(R) Streaming SIMD Extensions 4.2 (Intel(R) SSE4.2) instructions.
The processor must support the Intel(R) Advanced Vector Extensions (Intel(R) AVX) instructions.

This is due to the Intel mkl package getting picked up when you try to install something that depends on it-- Rosetta can translate some Intel instructions but not the specialized ones here. To avoid this, make sure to use the environment variable CONDA_SUBDIR=osx-arm64, which restricts the Conda environment to only use ARM packages, and use nomkl as described above.