Numpy v2.0.0 breaks the ability to download models using spaCy

[afogel]

How to reproduce the behaviour

In my dockerfile, I run these commands:

FROM --platform=linux/amd64 python:3.12.4

RUN pip install --upgrade pip

RUN pip install torch --index-url https://download.pytorch.org/whl/cpu
RUN pip install spacy

RUN python -m spacy download en_core_web_lg

It returns the following error (and stacktrace):

2.519 Traceback (most recent call last):
2.519   File "<frozen runpy>", line 189, in _run_module_as_main
2.519   File "<frozen runpy>", line 148, in _get_module_details
2.519   File "<frozen runpy>", line 112, in _get_module_details
2.519   File "/usr/local/lib/python3.12/site-packages/spacy/__init__.py", line 6, in <module>
2.521     from .errors import setup_default_warnings
2.522   File "/usr/local/lib/python3.12/site-packages/spacy/errors.py", line 3, in <module>
2.522     from .compat import Literal
2.522   File "/usr/local/lib/python3.12/site-packages/spacy/compat.py", line 39, in <module>
2.522     from thinc.api import Optimizer  # noqa: F401
2.522     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2.522   File "/usr/local/lib/python3.12/site-packages/thinc/api.py", line 1, in <module>
2.522     from .backends import (
2.522   File "/usr/local/lib/python3.12/site-packages/thinc/backends/__init__.py", line 17, in <module>
2.522     from .cupy_ops import CupyOps
2.522   File "/usr/local/lib/python3.12/site-packages/thinc/backends/cupy_ops.py", line 16, in <module>
2.522     from .numpy_ops import NumpyOps
2.522   File "thinc/backends/numpy_ops.pyx", line 1, in init thinc.backends.numpy_ops
2.524 ValueError: numpy.dtype size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject

Locking to the previous version of numpy will resolve this issue:

FROM --platform=linux/amd64 python:3.12.4

RUN pip install --upgrade pip

RUN pip install torch --index-url https://download.pytorch.org/whl/cpu
RUN pip install numpy==1.26.4 spacy

RUN python -m spacy download en_core_web_lg

[gborodin]

+1

[rustammdev]

How to reproduce the behaviour

In my dockerfile, I run these commands:

FROM --platform=linux/amd64 python:3.12.4

RUN pip install --upgrade pip

RUN pip install torch --index-url https://download.pytorch.org/whl/cpu
RUN pip install spacy

RUN python -m spacy download en_core_web_lg

It returns the following error (and stacktrace):

2.519 Traceback (most recent call last):
2.519   File "<frozen runpy>", line 189, in _run_module_as_main
2.519   File "<frozen runpy>", line 148, in _get_module_details
2.519   File "<frozen runpy>", line 112, in _get_module_details
2.519   File "/usr/local/lib/python3.12/site-packages/spacy/__init__.py", line 6, in <module>
2.521     from .errors import setup_default_warnings
2.522   File "/usr/local/lib/python3.12/site-packages/spacy/errors.py", line 3, in <module>
2.522     from .compat import Literal
2.522   File "/usr/local/lib/python3.12/site-packages/spacy/compat.py", line 39, in <module>
2.522     from thinc.api import Optimizer  # noqa: F401
2.522     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
2.522   File "/usr/local/lib/python3.12/site-packages/thinc/api.py", line 1, in <module>
2.522     from .backends import (
2.522   File "/usr/local/lib/python3.12/site-packages/thinc/backends/__init__.py", line 17, in <module>
2.522     from .cupy_ops import CupyOps
2.522   File "/usr/local/lib/python3.12/site-packages/thinc/backends/cupy_ops.py", line 16, in <module>
2.522     from .numpy_ops import NumpyOps
2.522   File "thinc/backends/numpy_ops.pyx", line 1, in init thinc.backends.numpy_ops
2.524 ValueError: numpy.dtype size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject

Locking to the previous version of numpy will resolve this issue:

FROM --platform=linux/amd64 python:3.12.4

RUN pip install --upgrade pip

RUN pip install torch --index-url https://download.pytorch.org/whl/cpu
RUN pip install numpy==1.26.4 spacy

RUN python -m spacy download en_core_web_lg

this solution helped, thank you

[supert56]

+1 I also had this problem. Thanks for posting the solution 👍

[nachthammer]

Those solutions indeed works, but I would still like to see a fix in the codebase itself. This issue is that inside the requirements.txt of the project (just an assumption after a short look at the codebase), the version is specified as such:

numpy>=1.15.0; python_version < "3.9"
numpy>=1.19.0; python_version >= "3.9"

I am a huge fan, in all of my projects, of always pinning dependencies even up to the patch version.

I would suggest a PR that looks like this:

numpy>=1.15.0,<2.0.0; python_version < "3.9"
numpy>=1.19.0,<2.0.0; python_version >= "3.9"

This at least pins the version down to major releases, which should anyway always be the case, as major version can (and most likely will always) contain breaking changes.

[afogel]

@DoctorManhattan123 To clarify, the solution I posted is only meant to be a stopgap.

Ideally, all downstream consumers of numpy (including library maintainers) should complete the migration to leverage numpy 2.0.0. I imagine, given the size of the release, that this will take time.

The pinned version is to tide over people seeking to quickly fix their CI/CD or whatever impacted process is broken until a more robust solution is implemented in the affected codebases.

[bendennescma]

This issue with thinc has been noted explosion/thinc#939

[lucas-mdsena]

It helped. Thanks!

[cyriaka90]

The new release 3.7.6 should resolve this :)

[ddayan]

I'm still experiencing the same error on 3.7.6 and numpy 2.1 && 2.0.0. As a sanity check it works after downgrading to 1.26.4

[CptCaptain]

The issue still persists with the 3.7.6-release as it still depends on thinc<8.3, which is incompatible with numpy>=2.0

[bendennescma]

The issue still persists with the 3.7.6-release as it still depends on thinc<8.3, which is incompatible with numpy>=2.0

Yes it appears thinc v8.3.0 itself is the first release that is compatible with numpy>=2.0

The latest release before that (v8.2.5) specifically restricts numpy pin to <2.0.0

[filbranden]

See also #13607

[honnibal]

Sorry for the delay on this.

I want to release the upgraded numpy pin as version 3.8, because I don't want to drop support for Python 3.8 in a patch release. Upgrading to numpy v2 in a patch release is also questionable.

However, the model artifacts bake in the version of spaCy into the package. This means I need to retrain the models to do the v3.8 release, and the retraining is taking some time.