Running Stable Diffusion on M1 MacBook Pro

Kyosuke Takayama — Wed, 24 Aug 2022 03:49:21 +0000

original article here: https://zenn.dev/ktakayama/articles/6c627e0956f32c

AI image generator Stable Diffusion is now open source. I want to running it on local machine but I only have a MacBook Pro then not to easy.

https://github.com/CompVis/stable-diffusion

The following thread is very helpful!

https://github.com/CompVis/stable-diffusion/issues/25

Speed

Here is my MacBook Pro 14 spec.

Apple M1 Pro chip
8 core CPU with 6 performance cores and 2 efficiency cores
14-core GPU
16-core Neural Engine
32GB memory

It needs about 15–20 GB of memory while generating images. 6 images can be generated in about 5 minutes.

Get model

https://huggingface.co/CompVis/stable-diffusion-v-1-4-original

Get source code

Get the source code in the apple-silicon-mps-support branch of this repository.

https://github.com/magnusviri/stable-diffusion/tree/apple-silicon-mps-support

Setup

Install conda and rust with homebrew.

brew install miniconda rust

Setup shell environment for conda. I use zsh.

conda init zsh

When I run conda env create, I get an error.

$ conda env create -f environment-mac.yaml
Collecting package metadata (repodata.json): done
Solving environment: failed

ResolvePackageNotFound:
  - python=3.8.5

Edit environment-mac.yaml to match your environment. Specifically, change the version number to match your environment. For example.

diff --git a/environment-mac.yaml b/environment-mac.yaml
index d923d56..c8a0a8e 100644
--- a/environment-mac.yaml
+++ b/environment-mac.yaml
@@ -3,14 +3,14 @@ channels:
   - pytorch
   - defaults
 dependencies:
-  - python=3.8.5
-  - pip=20.3
+  - python=3.9.12
+  - pip=21.2.4
   - pytorch=1.12.1
   - torchvision=0.13.1
   - numpy=1.19.2
   - pip:
     - albumentations==0.4.3
-    - opencv-python==4.1.2.30
+    - opencv-python>=4.1.2.30
     - pudb==2019.2
     - imageio==2.9.0
     - imageio-ffmpeg==0.4.2

activate and link to model.

conda activate ldm
mkdir -p models/ldm/stable-diffusion-v1
ln -s /path/to/stable-diffusion-v-1-4-original/sd-v1-4.ckpt models/ldm/stable-diffusion-v1/model.ckpt

Do image generation!

I get a PyTorch related error when I execute txt2image.

$ python scripts/txt2img.py --prompt "a photograph of an astronaut riding a horse" --plms 
〜 skip 〜
NotImplementedError: The operator 'aten::index.Tensor' 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. 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.

Install nightly version.

conda install pytorch torchvision torchaudio -c pytorch-nightly

This still got error.

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

Fix this error.

https://github.com/CompVis/stable-diffusion/issues/25#issuecomment-1221667017

vi /opt/homebrew/Caskroom/miniconda/base/envs/ldm/lib/python3.9/site-packages/torch/nn/functional.py

--- functional.py_      2022-08-23 17:07:29.000000000 +0900
+++ functional.py       2022-08-23 17:07:31.000000000 +0900
@@ -2506,9 +2506,9 @@ def layer_norm(
     """
     if has_torch_function_variadic(input, weight, bias):
         return handle_torch_function(
-            layer_norm, (input, weight, bias), input, normalized_shape, weight=weight, bias=bias, eps=eps
+            layer_norm, (input.contiguous(), weight, bias), input, normalized_shape, weight=weight, bias=bias, eps=eps
         )
-    return torch.layer_norm(input, normalized_shape, weight, bias, eps, torch.backends.cudnn.enabled)
+    return torch.layer_norm(input.contiguous(), normalized_shape, weight, bias, eps, torch.backends.cudnn.enabled)

Everything OK! Great!!!