Installation Instructions for Mac

Question

Installation Instructions for Mac

johnhorsema opened this issue 2 years ago · comments

Thanks for releaseing this stable diffusion model.
Was wondering if there is equivalent webui.cmd for Mac?

Timothy Strimple · Answer 1 · Thu Sep 01 2022 03:31:34 GMT+0800 (China Standard Time)

There isn't one included. You could create a bash script which would work for both Linux and Mac. If you look at the webui.cmd file it's setting up some environment variables and the conda environment before invoking python scripts/relauncher.py. I already had the environment setup, so I just ran relauncher.py directly.

Joshua Kimsey · Answer 2 · Thu Sep 01 2022 03:48:55 GMT+0800 (China Standard Time)

I already have the Linux Bash script made, but it was made explicitly with Linux in mind, not Mac. You'll have to do some work to make it work with a Mac. Further still, does Stable Diffusion even run correctly on an M1/M2 Mac? I thought that was still a planned feature sometime later on. Unless you have a Mac Pro, or an older Mac, which in that case you'll need to set it up to run with AMD hardware, which can be done.

Samson · Answer 3 · Thu Sep 01 2022 04:25:09 GMT+0800 (China Standard Time)

StableDiffusions itself works fine on Apple Silicon using the MPS backend.

Joshua Kimsey · Answer 4 · Thu Sep 01 2022 04:28:40 GMT+0800 (China Standard Time)

Interesting! What's the performance like on Apple Silicon? It is able to make use of their AI Co-Processors or is it just using it like a normal CPU to render?

Samson · Answer 5 · Thu Sep 01 2022 04:39:42 GMT+0800 (China Standard Time)

Happy to run a specific benchmark on my M1, but it takes O(minutes) to run txt2img using their AI Co-processors (vs 10's of minutes using just the cpu).

Rudolf Olah · Answer 6 · Thu Sep 01 2022 13:05:08 GMT+0800 (China Standard Time)

@fragmede did you have to modify the environment.yaml to remove the cudatoolkit package in order for the Conda environment creation to work? Was there another change you had to make to change the backend to use MPS?

Samson · Answer 7 · Thu Sep 01 2022 14:53:01 GMT+0800 (China Standard Time)

@rudolfolah I used environment-mac.yaml from https://github.com/fragmede/stable-diffusion/tree/mps_consistent_seed

Rudolf Olah · Answer 8 · Thu Sep 01 2022 20:41:02 GMT+0800 (China Standard Time)

@fragmede awesome, thanks for the link, I tried out the branch and it ended up working with this: PYTORCH_ENABLE_MPS_FALLBACK=1 python scripts/orig_scripts/txt2img.py --prompt "turtle drinking wine"

It took over 30min to generate 6 samples, seemed like a lot of ops fall back to cpu? It's still very cool to see progress, especially since PyTorch's official support for M1s started recently: https://pytorch.org/blog/introducing-accelerated-pytorch-training-on-mac/

Rudolf Olah · Answer 9 · Fri Sep 02 2022 03:20:29 GMT+0800 (China Standard Time)

someone posted instructions on their blog, which just made the frontpage of Hacker News: https://replicate.com/blog/run-stable-diffusion-on-m1-mac

the repo is here: ~~wrong repo~~ https://github.com/magnusviri/stable-diffusion

I haven't dug into it yet, so I'm not sure what changes they've made to ensure that MPS is used rather than CPU.

Eric · Answer 10 · Fri Sep 02 2022 12:32:02 GMT+0800 (China Standard Time)

I think most of the work for this is happening in lstein's fork: https://github.com/lstein/stable-diffusion/

the main differences are making sure to use the pytorch nightly, and ensuring that you're passing PYTORCH_ENABLE_MPS_FALLBACK=1 because not all PyTorch operators are implemented for the MPS backend yet. (which is just the Apple GPU backend, not the sweet NPU/dedicated Neural Engine). There also needs to be some logic implemented for choosing the backend device (which is actually the hard part, since there are a few files that need changing).

I don't know how well it would play with the additional GAN filters; sourcing macOS/Apple Silicon specific versions might be necessary. waifu2x might be able to help here.

I would love to see some of that work ported here; the WebUI is awesome.

Joshua Kimsey · Answer 11 · Fri Sep 02 2022 13:05:20 GMT+0800 (China Standard Time)

Imagine Stable Diffusion running on the Apple NPU (the thing I've been referring too as Apple's Co-Processor). That would be truly amazing too see. All the more so, given Emmad has talked about wanting to get the Model sizes down to a low as 100MB. That could open up doing stuff like this on mobile devices like the iPhone, or the Google Pixel 6 using it's Tensor Processors.

swyx.io · Answer 12 · Sun Sep 04 2022 17:27:27 GMT+0800 (China Standard Time)

just commenting to confirm that copying mac-environment.yml and then conda env create -f mac-environment.yaml worked for me. ( i also entered a venv first)

however i was unable to install gradio in conda, and all the gradio instructions are using pip install.

Rudolf Olah · Answer 13 · Tue Sep 06 2022 09:39:36 GMT+0800 (China Standard Time)

the scripts need to be updated to check for MPS vs CUDA, the environments-mac.yaml looks alright, I also saw the same issue with gradio not being installed when running the webui script. The txt2img script in the other repo was updated to check for MPS: invoke-ai/InvokeAI@bdb0651

Mathew Hartley · Answer 14 · Tue Sep 06 2022 10:56:56 GMT+0800 (China Standard Time)

just commenting to confirm that copying mac-environment.yml and then conda env create -f mac-environment.yaml worked for me. ( i also entered a venv first)

however i was unable to install gradio in conda, and all the gradio instructions are using pip install.

To solve this I added - defaults to the bottom of the channels section in the copied environment-mac.yaml from the other repo which then allowed gradio to install through conda.

As @rudolfolah said though, there's a lot of places which are hard-coded to cuda in this repo so it's a bit of work to switch it over at the moment.