numpy v2.0.2 fail: ValueError: numpy.dtype size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject

[rjurney]

NOTE: I figured out my issue. relik requires numpy==1.26.4 and it isn't pinned as a dependency in the package. This means your package manager is likely to install numpy==2.0.2 which is the latest as of January 30, 2025. This will confuse relik and it will not function. Downgrade to numpy==1.26.4 and all will be well...

My environment is:

• Anaconda Python 3.11.11
• conda environment
• Poetry 2.0.1
• relik==1.0.7
• Other than some dependencies like black and isort, relik is the only package I have installed via poetry add relik.

When I run the hello world:

from relik import Relik
from relik.inference.data.objects import RelikOutput

relik = Relik.from_pretrained("sapienzanlp/relik-entity-linking-large")
relik_out: RelikOutput = relik("Michael Jordan was one of the best players in the NBA.")

I get this numpy dimension exception, straight away:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[1], line 1
----> 1 from relik import Relik
      2 from relik.inference.data.objects import RelikOutput
      4 relik = Relik.from_pretrained("sapienzanlp/relik-entity-linking-large")

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/relik/__init__.py:1
----> 1 from relik.inference.annotator import Relik
      2 from pathlib import Path
      4 VERSION = {}  # type: ignore

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/relik/inference/annotator.py:26
     24 from relik.inference.data.splitters.base_sentence_splitter import BaseSentenceSplitter
     25 from relik.inference.data.splitters.blank_sentence_splitter import BlankSentenceSplitter
---> 26 from relik.inference.data.splitters.spacy_sentence_splitter import SpacySentenceSplitter
     27 from relik.inference.data.splitters.window_based_splitter import WindowSentenceSplitter
     28 from relik.inference.data.tokenizers.spacy_tokenizer import SpacyTokenizer

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/relik/inference/data/splitters/spacy_sentence_splitter.py:3
      1 from typing import Any, Iterable, List, Optional, Union
----> 3 import spacy
      5 from relik.inference.data.objects import Word
      6 from relik.inference.data.splitters.base_sentence_splitter import BaseSentenceSplitter

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/spacy/__init__.py:6
      3 from typing import Any, Dict, Iterable, Union
      5 # set library-specific custom warning handling before doing anything else
----> 6 from .errors import setup_default_warnings
      8 setup_default_warnings()  # noqa: E402
     10 # These are imported as part of the API

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/spacy/errors.py:3
      1 import warnings
----> 3 from .compat import Literal
      6 class ErrorsWithCodes(type):
      7     def __getattribute__(self, code):

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/spacy/compat.py:39
     36 except ImportError:
     37     from catalogue import _importlib_metadata as importlib_metadata  # type: ignore[no-redef]    # noqa: F401
---> 39 from thinc.api import Optimizer  # noqa: F401
     41 pickle = pickle
     42 copy_reg = copy_reg

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/thinc/api.py:1
----> 1 from .backends import (
      2     CupyOps,
      3     MPSOps,
      4     NumpyOps,
      5     Ops,
      6     get_current_ops,
      7     get_ops,
      8     set_current_ops,
      9     set_gpu_allocator,
     10     use_ops,
     11     use_pytorch_for_gpu_memory,
     12     use_tensorflow_for_gpu_memory,
     13 )
     14 from .compat import enable_mxnet, enable_tensorflow, has_cupy
     15 from .config import Config, ConfigValidationError, registry

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/thinc/backends/__init__.py:17
     15 from ._cupy_allocators import cupy_pytorch_allocator, cupy_tensorflow_allocator
     16 from ._param_server import ParamServer
---> 17 from .cupy_ops import CupyOps
     18 from .mps_ops import MPSOps
     19 from .numpy_ops import NumpyOps

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/thinc/backends/cupy_ops.py:16
      6 from ..util import (
      7     is_cupy_array,
      8     is_mxnet_gpu_array,
   (...)
     13     torch2xp,
     14 )
     15 from . import _custom_kernels
---> 16 from .numpy_ops import NumpyOps
     17 from .ops import Ops
     20 @registry.ops("CupyOps")
     21 class CupyOps(Ops):

File ~/anaconda3/envs/iecompare/lib/python3.11/site-packages/thinc/backends/numpy_ops.pyx:1, in init thinc.backends.numpy_ops()

ValueError: numpy.dtype size changed, may indicate binary incompatibility. Expected 96 from C header, got 88 from PyObject

It looks like relik expects a certain version of numpy that isn't getting installed correctly when I poetry add relik.

Please advice... I am writing a blog post entitled Three New Information Extraction Models Compared about this project, comparing it with GLiREL and NuExtract. This stopped me straight away. I am going to test the other two and come back here and debug this bad boy. In the meantime maybe you will get back to me :)