🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/158647

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

❌ 38 New Failures, 1 Unrelated Failure

As of commit bc4e307 with merge base 01f66d0 ():

This comment was automatically generated by Dr. CI and updates every 15 minutes.

It's too early to drop numpy 1.26.4 support, that will have to stay for quite a while longer since a lot of ML/AI-relevant libraries only fairly recently did releases with numpy >=2.0 support (so this would narrow the window of compatible versions for all those libraries) or even for a few straggelrs aren't compatible at all yet (e.g., sagemaker-python-sdk).

@rgommers FYI, the last numpy 1.x version (1.26) will be EoL in 2 months:

• https://endoflife.date/numpy

This PR is still a draft. This will not be merged very soon. Additionally, PyTorch 2.8 features are frozen, and they should support NumPy 1.x.

End-of-life for a Python package doesn't really mean much (nor is that even a thing for most projects) - it stays working just fine, it's not like for a distro that security bug fixes stop and it becomes dangerous to use that version.

You have to think about this in terms of compatibility windows. And dropping numpy 1.26 in particular will shorten that window by a lot for many other packages, as you can see from version bumps in this PR. The current approach seems to be to keep the window as wide as possible, and you're switching it to be pretty aggressive here. E.g., support is removed for:

• numpy versions 1.20-1.26
• numba versions 0.49 - 0.59
• tensorboard versions 2.13 - 2.19

Some tightening there seems healthy because some versions are too old, but this probably swings a bit too far in the other direction.

@albanD is there a policy on this similar to the policy for CPython support?

We could at least fix some bugs discovered from this PR.

We don't have a strict policy on this no.
I 100% agree we should not drop numpy 1.26 any time soon. And we should keep testing on the oldest possible numpy version in CI.

That being said, I 100% agree @cyyever that we should improve our numpy 2.X support.
I do think that having a cpu shard on numpy 2.X would help there!