AMD card support?

[agronholm]

As an owner of a Radeon 7900 XTX, I'm wondering if this project could be made to support AMD cards too. The problem is the xformers dependency which does not support AMD cards. Does Open-Sora-Plan use any nVidia-only features?

[LinB203]

It do not need to install xformers. I think it is ok in AMD card, but I do have this env to test. You can test and report some info here.

[agronholm]

Why is it declared as a dependency then, here? Also why are all the dependencies pinned to specific versions, rather than the minimum version? If you want to provide a pinned "this worked for us" list of dependencies, then the usual procedure for that would be to provide a requirements.txt which pins the versions of direct and transitive dependencies.

I also kind of wonder why the project needs something like pydub as a dependency, when this project does not even claim to produce audio?

[agronholm]

It looks like you already have a proper requirements.txt, here. So you don't need to pin exact versions in pyproject.toml too.

[agronholm]

Digging deeper, it looks like you just dumped everything currently in your virtualenv onto the dependencies list, instead of listing the libraries OpenSora actually needs. I'm trying to untangle the mess, will keep you posted.

[agronholm]

From a cursory search, these are the third party libraries actually directly imported by OpenSora:

• accelerate
• altair
• basicsr
• bs4
• colorlog
• decord
• deepspeed
• diffusers
• einops
• ftfy
• gdown
• gradio
• imageio
• lcalib (NOT declared as a dependency, NOT on the PyPI, can't even find it on Google!)
• mmcv (NOT declared as a dependency!)
• numpy
• omegaconf
• opencv-python
• packaging
• peft
• pillow
• tensorboard
• timm
• transformers
• torch
• torchvision
• torch_npu
• tqdm
• wandb

Additionally, many of these imports are only needed for training, not inference, so they should be moved to an extra, so as not to make users install dependencies they don't need.