Run locally on desktop rather than online colab?

[cragginstylie]

I've tried running this locally on a Windows 11 desktop with Nvidia 3060/12GB. It installs ok, but when I try the "Run locally" python script described in the README, I get the following error:

AssertionError: Torch not compiled with CUDA enabled

I'm running Auto1111 with no issues, and I'm confident that torch must be compiled with CUDA support there as well. Maybe the requirements aren't defined correctly to insure the correct binary for Torch is installed? I did the 'pip install -U stable_diffusion_videos` as directed...

I'm also hoping that this repo could be installed, similar to Auto1111, with a 'git clone' or 'git pull', followed by having a launch script (.bat and/or .sh) rather than going the package route.

Any tips would be appreciated. Also willing to help to make my suggestions possible, just need some guidance.

[nateraw]

hmm. I'm not sure (I don't have a windows machine, and even if I set up a windows machine in CI here, it won't have a GPU)

I think as long as you have SOME environment with torch correctly installed and working with GPU, it should be fine (i.e. you get True when you run import torch followed by torch.cuda.is_available() within your env).

I haven't worked with GPU in a long time (maybe 3 years) so I am sorry I'm unable to help more 😢 . Please let me know if you resolve this, and any steps you had to take. perhaps we can use that to update some instructions for windows users. I would appreciate that a lot.

[nateraw]

haha all good :) happy to try and help. tldr it sounds like if you just run the python from that venv you created with automatic, and install stable-diffusion-videos within it, you should be good to go.

---

Now, with regards to your suggestions:

Letting upstream diffusers install torch is probably not the best way, IMHO, because it obviously doesn't work

I think you'll find that this is the standard and accepted approach when it comes to torch. No need to add complexity (I'm referring to bat files in the automatic repo) to the project here, when the real issue is upstream at torch.

I'm wondering out loud... if we were to add a line to the reqs.txt just above the diffusers entry with another entry for torch with the correct version & extensions (i.e. torch=1.12.1+cu113) then it'll probably fix it.

I realize you're spitballin, which I appreciate, but this approach would cause more issues than it would solve:

• It doesn't take into consideration folks might not be using it on CUDA devices (this repo supports MPS on mac for example, and hopefully TPU at some point).
• That would be a hard coded version of cuda, which we definitely can't assume
• It's a very strict torch dep, whereas the current is anything >1.4.0

The repo you are referring to ends up running launch.py which is basically just installing stuff in a venv. You can easily do this yourself 😄 (if you don't know how, I PROMISE its worth taking a couple hours to dive deep into venv).

[nateraw]

Even better would be to use an installer/launch script similar to Auto1111 to set up a python venv with which to isolate all the package deps for better compartmentalization.

I'm incredibly uncomfortable with doing this, and really would prefer not to, but could be convinced if:

a. Someone with a windows machine does it (not me)
b. it doesn't add too much unnecessary complexity to the repo

[cragginstylie]

The repo you are referring to ends up running launch.py which is basically just installing stuff in a venv. You can easily do this yourself 😄 (if you don't know how, I PROMISE its worth taking a couple hours to dive deep into venv).

I know python well, and use venv frequently; however, mostly on linux - not windows. I'm using this windows-based gaming tower as my past experience was always that the newest video card drivers were often unavailable or unstable on linux, especially 3D drivers. With linux so much more mainstream now, I probably could've wiped windows 11, installed the latest Ubuntu and had no problems. I digress...

I'm incredibly uncomfortable with doing this, and really would prefer not to, but could be convinced if:

Well.....
a. I have a windows machine and can do it
b. which would be your call as to whether it adds too much complexity or not

Using python's venv is always a best practice, IMO. I assume this is a big reason you stated:

I PROMISE its worth taking a couple hours to dive deep into venv

LOL

Seriously, tho... I'll take a stab at it.

[nateraw]

haha really give the venv thing a try and let me know how it goes...then we can go from there.

I seriously bet you don't even need a new one if you already have the existing automatic111 one. Probably as simple as a call to the venv/.../Python.exe -m pip install stable-diffusion-videos then run whatever script you write with the library using the same executable... venv/.../Python.exe your_script.py. (the ... here means I don't know what the exact path would be).

[cragginstylie]

Exactly my plan. Thanks again for everything, especially your time!

[Raketemensch]

I'm not sure if you got this working, but I was able to get pytorch installed using Cuda 11.7 -- I don't think there's a version of pytorch that uses cuda 12 yet.

After pip installing stable-diffusion-videos (and the separate cuda 11.7 install) I then had to:

pip uninstall torch torchvision torchaudio
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117

This will install versions of the torch wheels built for cuda 11.7. Everything is working locally in Windows 10 for me now. I'd rather run it in Linux, but the good video card will sadly always be in a Windows machine...

[UnconnectedBedna]

Sorry for bump of necro thread.

What version of python is recommended for the pyenv?
I do not want to use the same as my stable diffusion but rather an isolated pyenv version.
(Idc about bloating my system)

My SD is running under 3.10.6, not the newest but it is what they recommend.