ci: Adding tests for numpy>=2

[KrishnanPrash]

What does the PR do?

Modifying L0_backend_python/examples/test.sh to test the python backend when numpy>=2 is installed in the environment. This is achieved by:

• Installing updated versions of numpy>=2.0, torch==2.5.0, and torchvision==0.20.0 at the start of the subtest
• Running through the existing subtest in L0_backend_python/examples/test.sh
• Uninstalling numpy>=2 and re-installing numpy<2 so future subtests will still use numpy 1.x

Reason for limited scope:
Support for numpy>=2 proved to be more complicated than intended because older version of packages like torch, torchvision, and tensorflow do not support for numpy 2.x.

Related PRs:

Python Backend PR: Link

Where should the reviewer start?

• CI Pipeline ID: 19960045

[rmccorm4]

Note that the if condition installed a +cu version, and the else did not. In your change you have both if/else installing the same thing. Please either match the install pattern with the correct +cuXXX variant, or double check whether this if/else is required if not. CC @fpetrini15 who might know a bit more about the if/else split here.

[fpetrini15]

@rmccorm4 I think this is a code segment we discussed a few months back. It seems like since Jacky has fixed my original error here; 23a6c21. I think for non-Jetson Linux tests, we want to continue installing the cuda-enabled version. That said, these tests were passing fine even with my logical error from months ago, so it's unclear to me what we achieve by installing the cuda-enabled version.

[rmccorm4]

If this doesn't break windows/jetson, I'm happy to keep it. If it might, we could also try making the change only on the linux non-jetson block. ex:

if not jetson/windows:
  pip install numpy 2
  pip install torch 2.5
else: # if windows/jetson
  pip install numpy 1
  pip install torch 2.0
fi

Let me know if you have a preference for minimal blast radius @fpetrini15

[fpetrini15]

I think we should align with the original intention of the script and install the cuda-enabled version.

CC @Tabrizian if you have any insight into why we preferred it originally?

[Tabrizian]

I believe PyTorch with CUDA was not supported on Jetson. We also don't support GPU tensors on Jetson with Python backend so it could be related to that.

[KrishnanPrash]

In L0_backend_python, the way examples/test.sh is called is along the lines of:

    setup_virtualenv

    set +e
    (cd ${TEST} && bash -ex test.sh)

    if [ $? -ne 0 ]; then
        echo "Subtest ${TEST} FAILED"
        RET=1
    fi
    set -e

    deactivate_virtualenv

So re-installing a lower version of numpy is unnecessary here because the virtualenv created for the examples subtest is not re-used for future subtests.

[rmccorm4]

LGTM but let's make sure the variants like python3.8, etc don't fail from this change. If they do, we might need to better handle which torch version is installed for each python version.

[krishung5]

LGTM. +1 on adding the comment for python 3.8.