Docker image: transformers-all-latest-gpu not running

Intermediate
1

I try to built my own docker container based on this image:

https://hub.docker.com/r/huggingface/transformers-pytorch-gpu/tags

which is probably sourced in this github repo:

https://github.com/huggingface/transformers/tree/main/docker/transformers-all-latest-gpu

That’s how I implement it, simple and clean:

FROM huggingface/transformers-pytorch-gpu:latest

This would first lead to the error docker no match for platform in manifest so I found out to provide a platform:

FROM --platform=linux/x86_64 huggingface/transformers-pytorch-gpu:latest

This will work, but when I eventually run the image I get two errors:


WARNING: The requested image's platform (linux/amd64) does not match the detected host platform (linux/arm64/v8) and no specific platform was requested
...
*************************
** DEPRECATION NOTICE! **
*************************
THIS IMAGE IS DEPRECATED and is scheduled for DELETION.

Further investigation leads to the doc of the underlying container image from nvidia/cuda: documentation Docker and it says:

The “latest” tag for CUDA, CUDAGL, and OPENGL images has been deprecated on NGC and Docker Hub.

Got. And when I use this as the base image, it seem to work - without providing a platform as mentioned above:

FROM nvidia/cuda:12.3.2-cudnn9-devel-ubuntu22.04

So I wonder: Am I the only to whom this happens? Is there something I am missing or is the docker image actually a little behind? Is there something I am doing wrong here?

Also I wonder if there’s some kind of a more extended documentation on the docker images, what would be the best to use.

FWIWF: I try to built this image on a Mac (to later use it on a Cuda-compatible system). I can’t just install the NVIDA Cuda Toolkit, which would be required to built the image on my own, because MacOS seems to be not supported.

cheers

1 Like