Is arm64 supported for diffusers?

[ghost]

I am not talking about Apple silicon.

I have tried to install diffusers on arm systems (ampere), diffusers["torch"] seems to clobber the torch install (if using https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch) and returns back errors about missing cuda libraries. When not installing with torch, diffusers doesnt seem to use the local pytorch, it also OOMs because it seems to be running off the CPU.

When trying without nvidia container (and just a plain ubuntu container), it seems like diffusers cannot install properly for an arm64 build.

So, the questions:

• Is arm64 CUDA supported?
• Can I use nvcr pytorch with diffusers and diffusers still use CUDA? if so how?

[tolgacangoz]

Can I ask which container version are you using? Which versions missing Cuda libraries are they? Also, which docker command did you run?
When installing on just a plain Ubuntu container, what errors does it show?

[sayakpaul]

Is arm64 CUDA supported?

We haven't tested this yet.

[ghost]

After more testing, diffusers seems to work with arm64. The versions I used are below.

accelerate==0.28.0
diffusers==0.26.3
transformers==4.38.2

I was able to get it working with nvcr.io/nvidia/pytorch:24.02-py3 which has pytorch arm64 image.

[sayakpaul]

Thanks for investigating!