You are not logged in.

#1 2025-09-17 04:42:39

John_Rochard
Member
Registered: 2025-08-05
Posts: 5

Sora AI package?

Hope this is the correct spot to ask about it.  I was banging my head against the keyboard last weekend, trying to install Sora Open AI to Arch Linux.  Here are the requirements for Sora;

torch==2.4.0
torchvision==0.19.0
colossalai>=0.4.4
mmengine>=0.10.3
ftfy>=6.2.0 # for t5
accelerate>=0.29.2 # for t5
av==13.1.0 # for video loading
liger-kernel==0.5.2
pandas>=2.0.3
pandarallel>=1.6.5
openai>=1.52.2
wandb>=0.17.0
tensorboard>=2.14.0
pre-commit>=3.5.0
omegaconf>=2.3.0
pyarrow

Then of course you get more requirements from these listed.  Almost akin to opening Pandora's box and releasing all the world's misery on your OS.  I have created a clean Arch partition on one of my SSD to do the voodoo.  Anyone got some shortcuts to follow?  I did Google this, got a Windows person on YouTube installing Ubuntu on his WSL and I don't care for Ubuntu or WSL.  Seems like it drove the poor guy mad insane too.  This weekend I might end up following his tutorial after the Ubuntu installation, unless I can find a tutorial in Arch.

Thanks everyone who read it, and thanks to the moderator who relocates it to the proper tree.

Offline

#2 2025-09-17 09:11:26

Lone_Wolf
Administrator
From: Netherlands, Europe
Registered: 2005-10-04
Posts: 14,280

Re: Sora AI package?

Are you referring to OpenAI Sora or Open-Sora ?

In both cases you should start a PKGBUILD, I'll assume yoiu want Open-Sora .
It seems to be python based, so you should start by reading https://wiki.archlinux.org/title/Python … guidelines .

Moderator Note
Moving to Creating & Modifying Packages

Last edited by Lone_Wolf (2025-09-17 09:12:13)


Disliking systemd intensely, but not satisfied with alternatives so focusing on taming systemd.

clean chroot building not flexible enough ?
Try clean chroot manager by graysky

Offline

#3 2025-09-17 09:27:28

loqs
Member
Registered: 2014-03-06
Posts: 18,618

Re: Sora AI package?

As an alternative to creating a PKGBUILD for it you could install it in a venv. This would also avoid you having to adjust the dependency versions to match those Arch currently packages.
Edit:
You may need to use `uv` and use python 3.12 if you want to avoid adjusting requirements.txt as python 3.13 support was only added in torch 2.5 while requirements.txt requires 2.4.

Last edited by loqs (2025-09-17 09:49:02)

Offline

#4 2025-09-17 10:18:18

Succulent of your garden
Member
From: Majestic kingdom of pot plants
Registered: 2024-02-29
Posts: 813

Re: Sora AI package?

For Open-Sora:

First you will need to create a virtual environment, probably if you are just starting to use python use pip, but I recommend uv after you checked that the virtual environment works. But first I will help you in pip  and conda so it can be more easy for you to do everything, since all the documentation of sora is in conda using pip.

So first create a virtual environment where you will work with open-sora as the repo says:

# create a virtual env and activate (conda as an example)
conda create -n opensora python=3.10
conda activate opensora

You will need to download pytorch, as the github repo says: you need to download it from here: https://pytorch.org/get-started/locally/  Select stable build, Linux, pip. We need to know which gpu do you have, if it is nvidia or amd. If it is amd then use the ROCm, then inside the virtual environment, it needs to be downloaded inside the virtual environment, so do it just after setup the virt env in the same terminal that you had executed the previous commands.

We need to know which gpu have because you will need to setup drivers and some packages to work. Specially in amd since you will need to install HIP to work well with pytorch.  https://wiki.archlinux.org/title/AMDGPU    https://wiki.archlinux.org/title/NVIDIA  If you are newbie to this things please wait a little after you tell us which gpu do  you have, probably because we can help you in choosing the right packages that you need.

then do the last part of the installation:

# download the repo
git clone https://github.com/hpcaitech/Open-Sora
cd Open-Sora

# Ensure torch >= 2.4.0
pip install -v . # for development mode, `pip install -v -e .`
pip install xformers==0.0.27.post2 --index-url https://download.pytorch.org/whl/cu121 # install xformers according to your cuda version
pip install flash-attn --no-build-isolation

you noticed that this line needs to be installed according to your cuda version or if you are using ROCm, check that before installing it, you will need to tweak it a little, probably just the url path.

pip install xformers==0.0.27.post2 --index-url https://download.pytorch.org/whl/cu121 # install xformers according to your cuda version

But I found the ofifclal repo https://github.com/facebookresearch/xformers Select one of the pips above according to your needs:

# [linux & win] cuda 12.6 version
pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu126
# [linux & win] cuda 12.8 version
pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu128
# [linux & win] cuda 12.9 version
pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu129
# [linux only] (EXPERIMENTAL) rocm 6.4 version
pip3 install -U xformers --index-url https://download.pytorch.org/whl/rocm6.4

then finally download the model:

pip install "huggingface_hub[cli]"
huggingface-cli download hpcai-tech/Open-Sora-v2 --local-dir ./ckpts

The installation of open-sora is very not friendly to my honest and humble opinion in comparison to other models.  Maybe you can ask if they can upload to ollama in github ? Maybe they are working on it, and you can just wait a little until model can be downloaded using ollama. But probably there is a reason why the model needs some libraries like xformers. In any case you can't loose nothing for just asking that. Probably the devs if there are interesting in their project, they are working in the ollama backend support. Then you can just download any Open UI and use the model.

Last edited by Succulent of your garden (2025-09-17 10:19:41)


str( @soyg ) == str( @potplant ) btw!

Offline

#5 2025-09-17 13:33:05

loqs
Member
Registered: 2014-03-06
Posts: 18,618

Re: Sora AI package?

Does ollama support text to video or image to video?  If it does not comfyu and diffusers both support a range of such models.

Offline

#6 2025-09-17 14:31:20

Succulent of your garden
Member
From: Majestic kingdom of pot plants
Registered: 2024-02-29
Posts: 813

Re: Sora AI package?

My bad you are right. it doesn't support it yet, It mostly LLMs or image to text models. But I found this links also, which seems promising in the future:

https://ollama.com/blog/vision-models
https://ollama.com/brxce/stable-diffusi … -generator

In some way maybe it can be made. But it's not the main feature or Ollama. I believed  that Ollama has already implemented stable diffusion models by some reason.  So probably Comfy UI or this one https://github.com/AUTOMATIC1111/stable-diffusion-webui which I setup to a friend a couple of years ago into his computer can be used as a frontend  [I'm not an user of image or videos generators, but I can help in setup the backend and running the models].


So it seems the only way to make a backend that runs sora is by following the github readme.

Last edited by Succulent of your garden (2025-09-17 14:32:31)


str( @soyg ) == str( @potplant ) btw!

Offline

#7 2025-09-18 02:55:15

killertofus
Member
Registered: 2025-02-10
Posts: 35

Re: Sora AI package?

pkgname=Open-Sora
pkgver=1.3
pkgrel=1
pkgdesc="Democratizing Efficient Video Production for All"
arch=(any)
url="https://github.com/hpcaitech/Open-Sora"
license=(Apache-2.0)
depends=('python' 'python-pytorch' 'python-torchvision' 'python-mmengine' 'liger-kernel' 'python-ftfy' 'python-accelerate' 'python-av' 'python-pandas' 'python-pandarallel' 'python-openai' 'python-wandb' 'python-tensorflow' 'pre-commit' 'python-omegaconf''python-pyarrow')
makedepends=('python-build' 'python-installer' 'python-wheel' 'python-setuptools-scm')
source=("$pkgname-$pkgver.tar.gz"::"$url/archive/v$pkgver.tar.gz")
build() {
  cd "$pkgname-$pkgver"
  python -m build --wheel --no-isolation
}
package() {
  cd "$srcdir/$pkgname-$pkgver"
  python -m installer --destdir="$pkgdir" dist/*.whl
}
sha256sums=('')



pkgname=colossalai
pkgver=0.5.0
pkgrel=1
pkgdesc="Making large AI models cheaper, faster, and more accessible"
arch=(any)
url="https://github.com/hpcaitech/ColossalAI"
license=(Apache-2.0)
depends=('python' 'python-pytorch 'cuda')
makedepends=('python-build' 'python-installer' 'python-wheel' 'python-setuptools-scm')
source=("https://files.pythonhosted.org/packages/source/${pkgname:0:1}/$pkgname/$pkgname-$pkgver.tar.gz")
build() {
  cd "$pkgname-$pkgver"
  python -m build --wheel --no-isolation
}
package() {
  cd "$srcdir/$pkgname-$pkgver"
  python -m installer --destdir="$pkgdir" dist/*.whl
}
sha256sums=('')




pkgname=liger-kernel
_pkgname=liger_kernel
pkgver=0.6.2
pkgrel=1
pkgdesc="Efficient Triton Kernels for LLM Training."
arch=(any)
url="https://github.com/linkedin/Liger-Kernel/"
license=(BSD-2-Clause)
depends=('python' 'python-pytorch' 'python-triton')
makedepends=('python-build' 'python-installer' 'python-wheel' 'python-setuptools-scm')
source=("https://files.pythonhosted.org/packages/source/${_pkgname:0:1}/$_pkgname/$_pkgname-$pkgver.tar.gz"
  "LICENSE")
build() {
  cd "$pkgname-$pkgver"
  python -m build --wheel --no-isolation
}
package() {
  cd "$srcdir/$pkgname-$pkgver"
  python -m installer --destdir="$pkgdir" dist/*.whl
}
sha256sums=(')


of course you will need to edit the sha256sums

Last edited by killertofus (2025-09-18 02:58:38)

Offline

#8 2025-09-19 04:35:51

John_Rochard
Member
Registered: 2025-08-05
Posts: 5

Re: Sora AI package?

Thanks to everyone who replied so quickly.

Lone_Wolf wrote:

I'll assume yoiu want Open-Sora .

Thank you for putting this subject where it belongs.  You assumed correctly.  I want to study some open source video AI and Open-Sora seems to be the correct choice.
Way back in the 90s about 30 some odd years ago, I use to compile POV-Ray models using scripts.  On the PCs back then it took the better part of a day to run one frame of a photo realistic model.  I am very patient when it comes to output.
They tell me one 4 second video at low resolution will take all day using this AI.  I know there are the web browser apps willing to charge fees or limit your time.  I want to experiment on my PC and not outsource it.

I won't have time today to hobby with this, but I do over Saturday and Sunday.  For reference purposes this video tutorial was all I could find on this subject; Open-Sora Installation Tutorial This person, however uses the Linux on Windows approach, I just don't care for at all. 

I will get back with you on Saturday and let you know my results.

Offline

#9 2025-09-19 20:19:31

loqs
Member
Registered: 2014-03-06
Posts: 18,618

Re: Sora AI package?

John_Rochard wrote:

want to study some open source video AI and Open-Sora seems to be the correct choice.

Is Wan2.2 too resource intensive for your system?

Offline

#10 2025-09-20 10:29:11

Lone_Wolf
Administrator
From: Netherlands, Europe
Registered: 2005-10-04
Posts: 14,280

Re: Sora AI package?

John_Rochard wrote:

For reference purposes this video tutorial was all I could find on this subject; Open-Sora Installation Tutorial This person, however uses the Linux on Windows approach, I just don't care for at all.

https://github.com/hpcaitech/Open-Sora#installation and https://deepwiki.com/hpcaitech/Open-Sor … -and-setup may be helpful.


Disliking systemd intensely, but not satisfied with alternatives so focusing on taming systemd.

clean chroot building not flexible enough ?
Try clean chroot manager by graysky

Offline

Board footer

Powered by FluxBB