Unexplained `numpy>=1.24.2` pin by `pytensor-base==2.10.1`

[michaelosthege]

Solution to issue cannot be found in the documentation.

• [X] I checked the documentation.

Issue

Trying to update to pytensor-base==2.10.1 (for PyMC 5.1.1) I'm stuck because pytensor-base==2.10.1 appears to require numpy>=1.24.2 but I can't figure out why. The recipe in this feedstock only requires numpy>=1.17.0.

mamba create -n pmtest -c conda-forge --dry-run "python==3.9.15" "numpy==1.22.4" "pytensor-base==2.10.1"

gives

Encountered problems while solving:
  - package pytensor-base-2.10.1-py39h99910a6_0 requires numpy >=1.24.2,<2.0a0, but none of the providers can be installed

This is very annoying because support for bleeding-edge NumPy is only available with very recent numba versions that no longer (or not yet) support Python 3.9.

Installed packages

does not apply

Environment info

does not apply

[maresb]

You're on osx-arm64?

[maresb]

No, windows

[maresb]

In https://anaconda.org/conda-forge/pytensor-base/2.10.1/download/win-64/pytensor-base-2.10.1-py39h99910a6_0.conda → info-pytensor-base-2.10.1-py39h99910a6_0.tar.zst → index.json:

{
  "arch": "x86_64",
  "build": "py39h99910a6_0",
  "build_number": 0,
  "depends": [
    "cons",
    "etuples",
    "filelock",
    "logical-unification",
    "minikanren",
    "numpy >=1.17.0",
    "numpy >=1.24.2,<2.0a0",
    "python >=3.9,<3.10.0a0",
    "python_abi 3.9.* *_cp39",
    "scipy >=0.14",
    "setuptools >=48.0.0",
    "typing_extensions",
    "ucrt >=10.0.20348.0",
    "vc >=14.2,<15",
    "vs2015_runtime >=14.29.30139"
  ],
  "license": "BSD-3-Clause",
  "name": "pytensor-base",
  "platform": "win",
  "subdir": "win-64",
  "timestamp": 1676935903169,
  "version": "2.10.1"
}

so indeed there is a requirement being added. I suspect it's 1503648 from https://github.com/conda-forge/pytensor-suite-feedstock/pull/16.

I'm guessing that it's the host dependency, so perhaps leaving it as a build and run dependency is sufficient?

Do you know where is that issue regarding the ABI compatibility from around the same time?

[maresb]

I'm surprised that a numpy pin is being added despite not using pin_compatible("numpy") as per https://conda-forge.org/docs/maintainer/knowledge_base.html#cross-compilation

[michaelosthege]

I got the problem in a Docker image build, so Linux. I tested the MWE on Windows, yes, but I have a strong prior that this is in fact independent of the OS.

[michaelosthege]

I also tried to inspect the .conda file, but couldn't open the .tar.zst inside. How did you do that?

[maresb]

I'm running Ubuntu and the default Archive Manager opens it no problem.

[hmaarrfk]

@conda-forge-admin please rerender

[conda-forge-webservices[bot]]

Hi! This is the friendly automated conda-forge-webservice.

I just wanted to let you know that I started rerendering the recipe in conda-forge/pytensor-suite-feedstock#35.