torch 2.x image does not contain dependencies according to requirements/*.txt

[ywchan2005]

🐛 Bug

In .azure/gpu-integrations.yml, two different images are used for torch 1.x and 2.x

• torch 1.x: pytorchlightning/torchmetrics:ubuntu22.04-cuda11.8.0-py3.9-torch1.13
• torch 2.x: pytorch/pytorch:2.1.1-cuda12.1-cudnn8-runtime

The image used in torch 1.x is built from dockers/ubuntu-cuda/Dockerfile and all required dev packages listed in requirements/_devel.txt are installed in advance

However, the image used in torch 2.x is a vanilla pytorch image which contains no dev packages required to build and test torchmetrics. As a result, even if a package is added in the corresponding .txt file in requirements directory, the image does not pick up the change and fails the build and check

To Reproduce

In my case,

1. add a new package in a .txt file under requirements directory
2. update the pull request to trigger the integration check which depends on torch 2.x image

Code sample

e.g. in requirements/image.txt, add kornia >=0.6.7, <0.7.1

Expected behavior

the images install packages according to the .txt files in requirements directory

Environment

• TorchMetrics version (and how you installed TM, e.g. conda, pip, build from source): 1.3.0dev
• Python & PyTorch Version (e.g., 1.0): 3.9.18 / 2.1.1+cu121
• Any other relevant information such as OS (e.g., Linux): Debian GNU/Linux 12 (bookworm)

Additional context

[github-actions]

Hi! thanks for your contribution!, great first issue!

[Borda]

@ywchan2005 thank you for your question; for integration, we would like to use only images with vanilla pytorch. Still, we faced recently breaking changes on Azure DevOps, such as any image used in workflows has to support node 20+, which means Ubuntu 20+, and the needed torch torch1.13 image is based on legacy Ubuntu 18, so we had to replace it...

If you need any rich docker images, please see https://hub.docker.com/r/pytorchlightning/torchmetrics/tags

[ywchan2005]

@Borda thanks for the reply. I'm in the middle of a pull request, and struggling how to move forward due to the errors raised in the integration pipeline which blocks the review process.

[Borda]

I'm in the middle of a pull request, and struggling how to move forward due to the errors raised in the integration pipeline which blocks the review process.

Have you already opened the PR? then we can see together where the problem is...

[ywchan2005]

Here's the PR
#2260