Ecosystem compatibility with numpy 2.0

[rgommers]

This list tracks the compatibility status of packages that depend on or support NumPy. If "compatible release on PyPI" does not say "yes" but a version number is listed: this is based on plans announced in a tracking issue or other communication by the authors of the package.

Maintainers: please feel free to edit directly (please refresh the page first to avoid overwriting edits from others!). Others who want to update things: please do comment, or feel free to ping me elsewhere.

Package name	Compatible release on PyPI?	Min compatible version	Notes
AstroPy	yes	6.1.0	astropy/astropy#16200
AwkwardArray	yes	2.6.3	scikit-hep/awkward#3064
BioPython		1.84	biopython/biopython#4676
Bokeh	yes	3.4.1	bokeh/bokeh#13835
Boost.Python		1.86.0	boostorg/python#431
Cartopy	yes	0.23	SciTools/cartopy#2339
ContourPy	yes	1.2.1	contourpy/contourpy#371
CuPy		14.0.0	cupy/cupy#8306
Cython	yes	3.0.4	Version is an estimate, it's worked fine for quite a while
Dask	yes	2024.5.1	dask/dask#11066
GDAL	yes	3.9.0	OSGeo/gdal#9751
GeoPandas	yes	0.14.4	geopandas/geopandas#3258
h5py	yes	3.11.0	h5py/h5py#2353
hypothesis	yes	6.100.2	HypothesisWorks/hypothesis#3950
imagecodecs			cgohlke/imagecodecs#100
Imageio			imageio/imageio#1077
JAX	yes	0.4.26	google/jax#19246
Keras			keras-team/keras#19691
Matplotlib	yes	3.8.4	matplotlib/matplotlib#26778
MDAnalysis			MDAnalysis/mdanalysis#4482
ml_dtypes	yes	0.4.0	jax-ml/ml_dtypes#143
netCDF4			Unidata/netcdf4-python#1317
NetworkX	yes	3.3	networkx/networkx#7390
Numba		0.60	(0.60.0rc1 is available) numba/numba#9544, Discourse post with context
numcodecs	yes	0.12.1	Likely older versions too; has been stable for a while. zarr-developers/numcodecs#521
numexpr	yes	2.10.0	pydata/numexpr#478
OpenCV			opencv/opencv-python#943
Pandas	yes	2.2.2	pandas-dev/pandas#55519
PyArrow	yes	16.0	apache/arrow#39532
Pybind11	yes	2.12.0	pybind/pybind11#5009
PyData Sparse	yes	0.15.2	pydata/sparse#680 (required Numba 0.60)
PyMC			depends on PyTensor
PySide6	yes	6.7.0	from a comment on this issue: #26191 (comment)
PyTables			PyTables/PyTables#1083
PyTensor			pymc-devs/pytensor#689
Pythran	yes	0.16.0	(0.15.0 works mostly, SciPy builds with it) serge-sans-paille/pythran#2189
PyTorch	yes	2.3.0	pytorch/pytorch#107302
PyWavelets	yes	1.6.0	PyWavelets/pywt#731
RAPIDS			rapidsai/build-planning#38
Rasterio	yes	1.3.10	rasterio/rasterio#3024
scikit-bio		0.6.1	scikit-bio/scikit-bio#1964
scikit-image	yes	0.23.1	scikit-image/scikit-image#7282
scikit-learn	yes	1.4.2	scikit-learn/scikit-learn#27075
SciPy	yes	1.13.0	scipy/scipy#20375
Shapely	yes	2.0.4	shapely/shapely#1972
SymPy		1.12.1
Seaborn	yes	0.13.2	mwaskom/seaborn#3683
statsmodels	yes	0.14.2	statsmodels/statsmodels#9194
TensorFlow		(maybe in May)	tensorflow/tensorflow#67291, has <2 upper bound for 2.16.1, requirements, lock file
threadpoolctl	yes	3.5.0	joblib/threadpoolctl#175
tifffile	yes	2024.4.24	cgohlke/tifffile#252
treelite	yes	4.2.1	dmlc/treelite#560
unyt	yes	3.0.2	yt-project/unyt#493
Xarray			pydata/xarray#8844
XGBoost		2.1.0	dmlc/xgboost#10221
yt	yes	4.3.1	yt-project/yt#4859
Zarr	yes	2.18.0	zarr-developers/zarr-python#1818
zfpy			LLNL/zfp#210

[jakevdp]

Thanks - for JAX we are planning an 0.4.26 release in the next day or two which will be built against NumPy 2.0.0rc1.

[jakevdp]

Also, in case you want to add it, we just released ml_dtypes v0.4.0 (https://pypi.org/project/ml-dtypes/) which is compatible with NumPy 2.0.

[jakirkham]

Thanks for putting this together Ralf! 🙏

This is incredibly helpful 🙂

[rgommers]

Thanks @jakevdp, I added the info for both JAX and ml_dtypes.

[neutrinoceros]

unyt 3.0.2 was also released a couple days ago for compat with numpy 2, in case you'd like to include it !

[neutrinoceros]

Also, here's where to track progress for cartopy : SciTools/cartopy#2339

[rgommers]

Thanks @neutrinoceros, much appreciated. Added both packages.

[ianthomas23]

ContourPy 1.2.1 has just been released on PyPI (https://pypi.org/project/contourpy/1.2.1/) with NumPy 2 compatibility, most relevant PR is contourpy/contourpy#371. It is a compulsory dependency of Matplotlib.

[jorisvandenbossche]

For PyArrow, it will be the upcoming 16.0 release that will be the first numpy-2.0-compatible release, but this will only be expected in around 3 weeks at the earliest (but normally certainly before the end of the month).
For people that need a PyArrow installed in an environment with numpy 2.0, it's worth noting that there are nightly wheels available though that already work with 2.0.

[dkbarn]

Could we add OpenCV to this list? I filed a ticket here to track progress on a numpy 2.0 build for it.

[rgommers]

Thanks @ianthomas23, @jorisvandenbossche, @dkbarn - all info added to the table.

[hawkinsp]

JAX released 0.4.26 on pypi, which is compatible with NumPy 2.0.

(NumPy folks: congratulations on your imminent v2 release!)

[ksunden]

mpl 3.8.4 is out, built with np 2

[neutrinoceros]

yt 4.3.1 is on PyPI and built with numpy 2.0.0rc1 !

EDIT(seberg): Added to table.

[neutrinoceros]

h5py 3.11 is here https://docs.h5py.org/en/stable/whatsnew/3.11.html#new-features

[h-vetinari]

Yes, Matti pointed out that patch to me (it wasn't linked in the issue so I didn't see it), and it works fine when backported to 1.84 and even 1.82.

[jakirkham]

@rgommers could we please update these entries in the table in the OP (replacing the old ones)?

Package name	Compatible release on PyPI?	Min compatible version	Notes
numcodecs	yes	0.12.1	Likely older versions too; has been stable for a while. zarr-developers/numcodecs#521
Zarr	yes	2.18.0	zarr-developers/zarr-python#1818

[rgommers]

Updated for all recent comments, thanks all!

[jakirkham]

Thanks Ralf! 🙏

Looks like nearly all libraries have some kind of issue/PR reference (or already a working release)

One that appears to be missed is Keras, so have raised upstream issue: keras-team/keras#19691

[jakirkham]

TensorFlow is targeting a branch cut end of May, we are unsure if numpy 2.0 support will make it but we are going to try and i should have better insight next week if the update will make next release. Like Jake said TensorFlow does have a <2.0 pin in at least both 2.16.1 and 2.15.1 which are our two supported versions currently, so I don't think we should immediately break if TensorFlow does not get the update done before a 2.0 release.

Thanks @MichaelHudgins for the insight here

Am curious how things are looking for TensorFlow + NumPy 2. Were you able to get more insight in the past 2 weeks?

Edit: Have gone ahead and opened an issue ( tensorflow/tensorflow#67291 ). Let's discuss there

[h-vetinari]

The boost version in the table is wrong. boostorg/python@0474de0 is not part of 1.85, but will be in 1.86.

PS. Sidenote: it would be helpful if the rest of the table (perhaps excluding some "most important packages") were sorted alphabetically, as it's getting a bit unwieldy to find specific packages.

[bryevdv]

it would be helpful if the rest of the table (perhaps excluding some "most important packages") were sorted alphabetically, as it's getting a bit unwieldy to find specific packages.

Agreed, if "important" packages need to be specially identified, they can be bolded or a column with an X or check mark or whatever

[bersbersbers]

I could not find an issue in their bug tracker, but some testing implies that PySide6 works from version 6.7.0 on:

# exits with "A module that was compiled using NumPy 1.x cannot be run in NumPy 2.0.0rc2 as it may crash."
uv pip install numpy==2.0.0rc2 PySide6-Essentials==6.6.3.1 && python -c "from PySide6 import QtWidgets"

# works fine
uv pip install numpy==2.0.0rc2 PySide6-Essentials==6.7.0   && python -c "from PySide6 import QtWidgets"

[jakirkham]

@rgommers have a few updates batched below for the table

GDAL 3.9.0 is released (with NumPy 2 support)! 🎉

• https://lists.osgeo.org/pipermail/gdal-dev/2024-May/058988.html

Also can we please add these NumPy 2 tracking issues to the table:

• TensorFlow: NumPy 2.0 support tensorflow/tensorflow#67291
• Keras: NumPy 2.0 support keras-team/keras#19691

At this point Numba's 0.60.0rc1 is packaged ( numba/numba#9544 (comment) ). Working on a conda-forge RC package for it as well. Would start pointing libraries downstream of Numba to these

Boost.Python should be fixed in the 1.86.0 release (we narrowly missed 1.85.0)

Dask is looking promising ( dask/dask#11066 (comment) ). Though still want to double check things there

Lastly opened an issue for Sparse: pydata/sparse#680

[rgommers]

Thanks all, updated for all comments.

PS. Sidenote: it would be helpful if the rest of the table (perhaps excluding some "most important packages") were sorted alphabetically, as it's getting a bit unwieldy to find specific packages.

Okay, I reordered it. The initial order has served its purpose, everything was blocked on a small set of packages at the start but those all did their release.

[hameerabbasi]

Lastly opened an issue for Sparse: pydata/sparse#680

sparse release is up on PyPI: https://pypi.org/project/sparse/0.15.2/

[jakirkham]

Dask 2024.5.1 works with NumPy 2

ref: https://docs.dask.org/en/stable/changelog.html#numpy-2-0-support

[rgommers]

Great, thanks @hameerabbasi and @jakirkham. Updated the table for those, and also added RAPIDS (linking to your tracking issue @jakirkham).

We're getting there - we've had ~7 weeks since RC1 and are about 75% through the list in this issue. In the community meeting next Wednesday we should probably discuss picking a release date in June.

[bnavigator]

What about (optional) inter-dependencies? Pandas and numexpr both claim to be compatible, but together they choke on the changed numpy 2 type repr of scalars: pandas-dev/pandas#58548

[rgommers]

I'd see that as a regular bug - a few test failures in a combo of two packages that were not tested together in CI is common, but doesn't detract (from the purposes of this tracking issue at least) from them both being NumPy 2.0-compatible.

[rgommers]

Posting here as well for completeness (xref #24300 (comment)): release date will be June 16th.

[peterjc]

For Biopython we do have C code using NumPy, but my understanding is we are insulated from the C API changes due to using the buffer protocol rather than the NumPy C API.

I expect a "fully compatible" Biopython 1.84 in early June, but all this addresses is the scalar repr changes to work on recent Numpy v1 and the v2 release candidates biopython/biopython#4676

https://lists.open-bio.org/pipermail/biopython/2024-May/017049.html

[jakirkham]

Updated the table for those, and also added RAPIDS (linking to your tracking issue @jakirkham).

Thanks Ralf! 🙏 Also thank you for including us 🙂

[hcho3]

Treelite 4.2.1 is compatible with NumPy 2.0