Compatibility with PyTorch >= 2.8 for Future Hardware (e.g., RTX 50 series)

[unumbrela]

Hi everyone,

I just wanted to share a successful workaround for installing MONAI on a system with an RTX 50 series GPU (5070 Ti), which requires a newer PyTorch version (2.8 preview in my case) than what is currently allowed by the pip package dependencies (torch<=2.6).

My Setup:

Ubuntu 24.04 (WSL2)
RTX 5070 Ti
CUDA 12.9
PyTorch 2.8.0 (preview)

The Workaround:

I downloaded the MONAI v1.5.0 source code from the releases page.
https://github.com/Project-MONAI/MONAI/releases

I ran pip install . from the source directory. This installed MONAI but also downgraded PyTorch to 2.6.

I immediately re-installed my required PyTorch 2.8.0 version.

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

After these steps, both MONAI 1.5.0 and PyTorch 2.8.0 are co-existing happily in my environment.

This seems to be a dependency constraint issue in setup.py. While my method is a bit of a hack, it does work. Hope this helps anyone else with next-gen hardware!

[pcanas]

It would be great if we could update monai in order to support Pytorch >= 2.6. Is this plan anytime soon?

[ericspod]

MONAI has now been updated to support 2.7. @unumbrela thanks for looking into this, if there's anything that needs to be changed to support 2.8 please do make a PR and we'll consider it. 2.8 will be released soon and we want to support it then, though we lack 50 series hardware for testing purposes.

[ogencoglu]

Any updates to support newer torch versions?

uv gives:

Because monai==1.5.0 depends on torch>=2.4.1,<2.7.0

[jamesobutler]

With cafc1fe adding PyTorch 2.8 support @ericspod will there soon be a new published monai version to PyPI to then be able to use PyTorch 2.7 and 2.8?

[ericspod]

Yes, we plan on a release soon to include support along with a number of fixes.

[garciadias]

I can confirm that installing the monai-weekly solves the problem and works well with uv installations.
I did not observe any issues, but I will run further tests today. monai-weekly

[garciadias]

Hi @ericspod,

I further tested this by using docker image nvcr.io/nvidia/pytorch:25.08-py3 .

To do that, I run the ./runtests.sh -u -q inside the container using the main branch and branch 6327a86, which I believe is the version deployed at monai-weekly (please confirm).

These are the results:

results_weekly.txt
results_main.txt
results_weekly_net.txt

Basically, the tests fail with monai 1.5.0, and most of the tests pass with monai-weekly. There are a couple of issues with the former as well, but I don't think they are critical.

Please, let me know if you want me to run any other tests.

[ericspod]

Thanks @garciadias! The hash for weekly was probably b5bc69d but yours is close enough. I think we're sure enough that it's working with 50 series now. We'll return the question of testing hardware later.

[garciadias]

Thanks @garciadias! The hash for weekly was probably b5bc69d but yours is close enough. I think we're sure enough that it's working with 50 series now. We'll return the question of testing hardware later.

Thank you, @ericspod . I will try that and use --net --coverage and let you know.

[garciadias]

I added the report to my previous comment. There are some errors, but nothing seems to be related to GPU, CUDA, or Torch.
I observed many tests running with substantial GPU load. So, yes, I think monai-weekly is working fine with the 50 series.