Xformers official wheel is available now

[aliencaocao]

https://github.com/facebookresearch/xformers/actions/workflows/wheels.yml
You can download them from github action artifacts
facebookresearch/xformers#588

No need to build from source yourself
Note that this only works if you have torch1.13+cu117 installed, which is NOT what the webui defaults to, but you can always install yourself to override it.
Works with python 3.7-3.10 on Windows and Linux

[bach777]

What are the improvements over the previous version?

[aliencaocao]

the current one that the webui installs is version 0.14.0dev which is 2 versions older than the current 0.16.0dev. It includes bug fixes and performance improvements that xformers repo committed

[bach777]

I see...I just need yo relaunch webui with --xformers argument in order to update?

[aliencaocao]

no, you need to download the wheel yourself and pip install from wheel, then launch with --xformers.

[hithereai]

On windows? None.

[ice051128]

https://github.com/facebookresearch/xformers/actions/workflows/wheels.yml You can download them from github action artifacts facebookresearch/xformers#588

No need to build from source yourself Note that this only works if you have torch1.13+cu117 installed, which is NOT what the webui defaults to, but you can always install yourself to override it. Works with python 3.7-3.10 on Windows and Linux

TLDR? What does this mean for people with --xformers that works out of the box without doing anything?

[bach777]

torch 1.12.1+cu113
This is the version I have installed, what I have to do to update? Please some help for this soul, I don't want to mess up my installation

[aliencaocao]

@ice051128 you need to enter the venv

[aliencaocao]

@bach777 pip3 install torch==1.13.0 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117
You need to do it in the venv if you are using 1. As for how to enter the venv, pls google yourself as it differs from system setup

[ice051128]

@bach777 pip3 install torch==1.13.0 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117 You need to do it in the venv if you are using 1. As for how to enter the venv, pls google yourself as it differs from system setup

Isn't torch 1.13.0 slower than 1.12.1 tho pytorch/pytorch#85297 (comment)

Can u use the new xformers with 1.13.1?

[aliencaocao]

I tried to, but I cannot build from source due to some issues. It should be either a bug on xformer side or my own environment's issue.
As for performance, that only concerns MPS, and xformers don't even support MPS, so there is no conflict here

[ataa]

I just installed the latest whl on Windows, now when I try to start the UI, I get:

A matching Triton is not available, some optimizations will not be enabled.
Error caught was: No module named 'triton'

Tried to install triton using pip install triton and pip install triton==2.0.0.dev20221120 and got this error:

ERROR: Could not find a version that satisfies the requirement triton (from versions: none)
ERROR: No matching distribution found for triton

How to fix this?

[aliencaocao]

Triton is not supported on windows.

[ataa]

Triton warning started to appear after upgrading to the latest Torch and Xformers whl. There was no warning prior to the upgrade.

[aliencaocao]

Yes because only xformer 0.15.0 onwards it uses triton

[ataa]

Thanks for your clarification.

[Mikerhinos]

Did we have to build it on Windows previously ?
Because I just start with argument --xformers and it seems it automatically installs and run it at launch.

[aliencaocao]

If you use torch 1.12.1+cu113, no. Anything else, yes.

[ice051128]

How do u revert the updates for torch and xformers? They have only caused random illegal memory access when generating images

[aliencaocao]

You can just use the pip command from offcial website for stable release. As for xformers, they have official wheel for 1.13.1

[ice051128]

You can just use the pip command from offcial website for stable release. As for xformers, they have official wheel for 1.13.1

Can't find the wheel for 1.13.1 anywhere. Is this the one?
They have a new wheel but not sure what it's built against
https://test.pypi.org/project/formers/0.0.16rc389/#files

[aliencaocao]

Sorry, not this one, you can see here https://github.com/facebookresearch/xformers/actions/runs/3765948564/jobs/6401935823

[hithereai]

Just tested torch 1.13.1 + cu117 on python 3.10.6, windows and 4090 + latest xformers 0.0.16rc396.
No speed change from the reg 1.12.0+cu113+xformers 0.0.14 pre-provided whl.

Still 25-29 it/s for a 70-steps 512/512 euler_a or LMS.

[hithereai]

Your whl is for p39 and I have 3.10.6 installed + couldn't get the latest .16 xformers to compile last night, will compile v15 or 14 and test.

Cheers.

[aliencaocao]

Right. Check this https://1drv.ms/u/s!AvJPuRJUdWx_8hi5p0ecOiV3yJtT?e=Nd6rE5
Build with 3.10 and torch 2.0 1224dev

[hithereai]

Just tested, getting 30-33 it/s compared to 25-29. Not bad!

[ClashSAN]

nice @hithereai , what flags are you using + batch size? because that's the highest it/s I've seen so far.

[hithereai]

@ClashSAN it's a fresh install of the latest commit (c6f347b) + --xformers flag + latest cudnn 8.7 in my torch/lib folder. Nothing else.

The reported speeds are for:
Batch size 1, pic size 512*512, 100 steps, samplers Euler_a or LMS.

With batch size 20 and 512*512 I still get the same total it/s as before - 45 (2.25 it/s per picture x 20).

Windows 11 22H2, 12700KF, stock settings 4090 FE.

Edit: also getting 25 it/s compared to 20-21 it/s for bs1 and 512/768 res.

[dragokas]

Where can I download the old xformers, intended for torch v.1.12.1?

[Nacurutu]

Where can I download the old xformers, intended for torch v.1.12.1?

https://github.com/C43H66N12O12S2/stable-diffusion-webui/releases/download/f/xformers-0.0.14.dev0-cp310-cp310-win_amd64.whl