Add support for Python 3.11, require NumPy 1.23+ #15111
Conversation
jameslamb
changed the title
Co-authored-by: Bradley Dice <bdice@bradleydice.com>
jameslamb
added
improvement
| @@ -65,6 +65,7 @@ requirements: | |||
| - scikit-build-core >=0.7.0 | |||
| - setuptools | |||
| - dlpack >=0.5,<0.6.0a0 | |||
| - numpy 1.23 | |||
There was a problem hiding this comment.
As NumPy is used at build time, this now adds it to requirements/host
| @@ -83,7 +84,7 @@ requirements: | |||
| - pandas >=2.0,<2.2.2dev0 | |||
| - cupy >=12.0.0 | |||
| - numba >=0.57 | |||
| - numpy >=1.21 | |||
| - {{ pin_compatible('numpy', max_pin='x') }} | |||
There was a problem hiding this comment.
Switch to using pin_compatible with NumPy to ensure the minimum version allowed at runtime matches what we built with and any NumPy version (pre-2.0.0) is allowed
| - output_types: pyproject | ||
| packages: | ||
| # Hard pin the patch version used during the build. | ||
| # Sync with conda build constraint & wheel run constraint. | ||
| - numpy==1.23.* |
There was a problem hiding this comment.
This is the same NumPy pin the conda packages use to build. It is now used here to align the wheel packages
| run_common: | ||
| common: | ||
| - output_types: [conda, requirements, pyproject] | ||
| packages: | ||
| - fsspec>=0.6.0 | ||
| - *numpy | ||
| - numpy>=1.23 |
There was a problem hiding this comment.
Both conda & wheel packages now have the same lower bound on NumPy
| @@ -625,7 +633,7 @@ dependencies: | |||
| - output_types: pyproject | |||
| packages: | |||
| - msgpack | |||
| - &tokenizers tokenizers==0.13.1 | |||
| - &tokenizers tokenizers==0.15.2 | |||
There was a problem hiding this comment.
I've bumped tokenizers to 0.15.2 so we can get Python 3.11 binaries.
There was a problem hiding this comment.
Incompatible with transformers==4.24.0 :(
There was a problem hiding this comment.
Bumped transformers to 4.38.1.
| tensor = torch.tensor(index) | ||
| got = cudf.Index(tensor) | ||
| assert_eq(got, index) | ||
| # TODO: This test fails with PyTorch 2. Is it still expected to be valid? |
There was a problem hiding this comment.
I pushed a change 2f0a8a8 to unblock CI, and also rolled it out to a separate PR #15188. See my notes there for discussion and next steps. https://github.com/rapidsai/cudf/pull/15188/files#r1506969204
I would like to merge this test change as-is, to keep the main goal of Python 3.11 support moving. We can handle any further discussion of the issue in https://github.com/rapidsai/cudf/pull/15188/files#r1506969204.
|
CI failed due to the dask tokenizing bug that was fixed in #15187. I will merge the upstream to rerun CI now. |
|
/merge |
|
Hooray! 🥳 Thanks everyone 👏 Glad to have this complete |
This PR contains a few minor improvements for `__cuda_array_interface__` and its tests. Found while working on #15111. Authors: - Bradley Dice (https://github.com/bdice) Approvers: - GALI PREM SAGAR (https://github.com/galipremsagar) URL: #15188
Description
Contributes to rapidsai/build-planning#3
This PR adds support for Python 3.11.
It also bumps uses of
NumPytonumpy>=1.23, see rapidsai/build-planning#3 (comment).Notes for Reviewers
This is part of ongoing work to add Python 3.11 support across RAPIDS.
The Python 3.11 CI workflows introduced in rapidsai/shared-workflows#176 are optional... they are not yet required to run successfully for PRs to be merged.
This PR can be merged once all jobs are running successfully (including the non-required jobs for Python 3.11). The CI logs should be verified that the jobs are building and testing with Python 3.11.
See rapidsai/shared-workflows#176 for more details.
(created with rapids-reviser)
Checklist