Ideas for speeding up with M1Mac

[autumnmotor]

I'm running stable-diffusion-webui on M1Mac (MacStudio 20coresCPU,48coresGPU, Apple M1 Ultra, 128GB RAM 1TB SSD).
It runs faster than the webui on my previous M1 Macmini (16GB RAM, 512 GB SSD), and I'm happy with it (although it was a little more expensive :D).

However, I would like to configure MacStudio to use its abundant resources more effectively and speed up processing even further.

It is a resource (load) image of the PC when continuously "Batch img2img" processes 384*640 size images (according to the application "iStat"), but the GPU is only using 55% and the CPU is only 11%. Even with 128GB of memory, python consumes about 10-16GB.

Is it possible to use these extra resources to do things faster?
Since I don't have much knowledge of computer architecture, I was searching for settings to speed things up at random. for example…

webui.py options
--opt-split-attention: slightly faster
--opt-channelslast: pretty slow
--disable-safe-unpickle: A bit faster (maybe?). But there are risks.
—use-cpu all: too slow.

Python library
Change pytorch from stable to Preview (Nightly). : Fairly fast.
xformers : Currently, it seems that you can't install it on Mac.

others
Placing stable-diffusion-webui on the ramdisk: Loading large file sizes is faster, but image processing itself remains the same.

It's really hit and miss. The best bet would be "best machine with best GPU", any other Mac ideas for speeding things up?

Translated by Google.
Thank you for reading my poor English.

[remixer-dec]

Can you please share your stats? How long does it take to generate a single 512x512 image at 20, 50 steps? Can you generate 2048x2048 images with that amount of RAM?
In my case, with 24 cores and 32gb, GPU is used at 80-90% and here are my stats
Also increasing "batch size" might help to use resources more efficiently, even though it will not increase the speed for a single image.

[autumnmotor]

txt2img(average on Batch count=10)
Euler a, 20 steps -> 7.8sec(2.45-2.71it/s)
Euler a, 50 steps -> 19.1sec(2.62-2.68it/s)

img2img(average on Batch count=10)
Euler a, 20 steps -> 6.2sec(2.45-2.71it/s)
Euler a, 50 steps -> 14.5sec(2.63-2.70it/s)

python: 10-16GB dedicated, GPU: 92-95%

2048x2048 image Impossible. Even 1024*1024 crashes with the error message below.

****/python3.10/multiprocessing/resource_tracker.py:224: UserWarning: resource_tracker: There appear to >be 1 leaked semaphore objects to clean up at shutdown
warnings.warn('resource_tracker: There appear to be %d '

Increasing the batch size is a good idea. Thank you.
I also think that this is certainly effective in making full use of resources.
I was able to confirm that 10-16GB exclusive use of python increased to 20,30GB, and I was trying to use RAM efficiently.
However, the image obtained by Batch size > 1 seems to have a lot of "noise" for some reason,
In some cases, even if Batch size=2, it may be forcibly terminated.

Also, increasing the batch size is a way to output multiple images from one input image.
As for me, in Batch img2img, I want to "output one image for one input image".
Because I want to change the touch and style of the video by prompt.
Therefore, mp4 is converted to a sequence image, Batch img2img is applied, and the output image group is converted to mp4 again (using ffmpeg etc.) as a sequential image.
In short, I want Batch img2img to execute "output one image for one input image" in parallel and perform processing efficiently (preferably at high speed).

By starting multiple webui.py, it was possible to process in parallel. However, this method slows down each webui.py and is (slightly) less efficient than a single run of webui.py.

It has become a long sentence without stopping...
Thanks for reading this far.

[bbecausereasonss]

Can you please provide instructions on this?

Change pytorch from stable to Preview (Nightly). : Fairly fast.

[kayzen-ml]

I'd be also glad to know how you did, @autumnmotor!

[gabrielbacha]

I just tried both --opt-split-attention and using torch nightly but I am still getting 2.05 it/s for a euler a 512x512 (unchanged vs before)