Add support for PyTorch 2.1

[shchur]

Should we increase the dependency ranges in _setup_utils.py to torch>=2.0,<2.2 and, potentially, lightning>=2.0,<2.2 in the AutoGluon v1.0 release?

To do:

• [ ] #3708
• [ ] #3827

@zhiqiangdon @tonyhoo @Innixma

[stevenmanton]

I'm interested in this. We want to use torch 2.1 but cannot alongside AutuGluon. The 2.1 restriction seems a bit too stringent. Was this done out of an abundance of caution or is there a known incompatibility?

[Ludecan]

It would be awesome if you guys could pull this out. I'm having issues with Neural Net models because of a bug in torch==2.0.1 (the one that gets installed by default with AutoGluon==1.0.0). Apparently torch didn't install the required CUDA libraries needed to work and it was fixed in 2.1.1: https://github.com/pytorch/pytorch/issues/100974

Is there any way to manually install these?

Attached is the error log I'm getting error.txt

Edit this is happening on an Ubuntu 24.04 WSL2 instance where my dependencies are handled via poetry.

[ddelange]

fixed by https://github.com/autogluon/autogluon/pull/3982