Docling depends on PyTorch which dropped support for macOS x86_64 after the 2.2.2 release. This old version of PyTorch works only with NumPy 1.x.

Trying to keep a simple installation for all Docling users, we tried to use special markers for it

'numpy (>=1.24.4,<3.0.0) ; sys_platform != "darwin" or platform_machine != "x86_64"',
'numpy (>=1.24.4,<2.0.0) ; sys_platform == "darwin" and platform_machine == "x86_64"',

Unfortunately, this creates non-negligible downstream issues when trying to add Docling to projects using dependency management systems like uv and poetry.

Old approach

All downstream project would be forced to adopt one of these approaches

1. Use exactly the same markers, or
2. Flag the project as non supporting macOS x86_64 systems, e.g. via

   [tool.uv]
   environments = ["sys_platform != 'darwin' or platform_machine != 'x86_64'"]

New approach

We are now lifting the extra markers for NumPy. As a consequence

1. Downstream projects and integrations will have much more freedom when including Docling
2. macOS x86_64 users must make sure to install a NumPy compatible version

In short, for macOS x86_64 install Docling as

pip install docling "numpy<2.0.0"