liblapack broken with numpy 2.1.3?

[jacobwilliams]

Solution to issue cannot be found in the documentation.

• I checked the documentation.

Issue

This environment.yml file worked a couple weeks ago, but does not work now:

channels:
  - nodefaults
  - conda-forge

dependencies:
  - mkl=2025.0.0
  - python=3.12.8
  - numpy=2.1.3

The error is:

Channels:
 - conda-forge
Platform: win-64
Collecting package metadata (repodata.json): done
Solving environment: failed

LibMambaUnsatisfiableError: Encountered problems while solving:
  - package numpy-2.1.3-py312h49bc9c5_0 requires liblapack >=3.9.0,<4.0a0, but none of the providers can be installed

Could not solve for environment specs
The following packages are incompatible
├─ mkl 2025.0.0**  is requested and can be installed;
├─ numpy 2.1.3**  is installable with the potential options
│  ├─ numpy 2.1.3 would require
│  │  └─ liblapack >=3.9.0,<4.0a0  but there are no viable options
│  │     ├─ liblapack 3.9.0 would require
│  │     │  └─ mkl <2025 , which conflicts with any installable versions previously reported;
│  │     └─ liblapack 3.9.0 would require
│  │        └─ mkl >=2024.2.2,<2025.0a0 , which conflicts with any installable versions previously reported;
│  ├─ numpy 2.1.3 would require
│  │  └─ python_abi 3.10.* *_cp310, which can be installed;
│  ├─ numpy 2.1.3 would require
│  │  └─ python_abi 3.11.* *_cp311, which can be installed;
│  ├─ numpy 2.1.3 would require
│  │  └─ python_abi 3.13.* *_cp313t, which can be installed;
│  └─ numpy 2.1.3 would require
│     └─ python_abi 3.13.* *_cp313, which can be installed;
└─ python 3.12.8**  is not installable because it requires
   └─ python_abi 3.12.* *_cp312, which conflicts with any installable versions previously reported.

It's hard to say if it's a liblapack problem but it seems that there have been many updates to the package since then? Based on looking at the file dates here: https://anaconda.org/conda-forge/liblapack/files

Did something break?

Installed packages

na

Environment info

na

[jacobwilliams]

More info. I have a working environment (created some weeks ago) with:

  - mkl=2025.0.0=hb752886_930
  - liblapack=3.9.0=8_h719fc58_netlib

When I try to create a new env with these packages, it doesn't work. How is that possible? Have these packages been retroactively changed?

[h-vetinari]

You cannot install mkl 2025 yet. It's broken currently, and we've started enforcing this constraint in #79 now.

[jacobwilliams]

But this worked for me for weeks. Did the old packages get changed? Why can't I create an env using the packages that worked before? It's very odd to me.

[h-vetinari]

MKL 2025 is incompatible with our blas infrastructure, and despite the fact that it was present in your environment, it was not actually being used by any of the packages that followed our regular blas setup. On the contrary, it lead to substantial problems where MKL-users would get switched to a different blas implementation (e.g. OpenBLAS), because of the way the blas variants interact with the solver, who's trying to maximize all versions in the environment: conda-forge/blas-feedstock#134

Why can't I create an env using the packages that worked before? It's very odd to me.

It didn't work as you expected, and caused persistent problems for other users, so indeed that combination was retroactively made impossible. I would have hoped the situation didn't get to this point, but Intel has taken a very long time to help fix the regressions in MKL 2025, and until they are fixed, you cannot use MKL 2025 together with the vast majority of our other packages, like numpy.

[jacobwilliams]

Ouch. Well this change certainly has broken my workflow. I'm using the 2025 intel ifx compiler and mkl for non-python reasons. It doesn't matter to me what lapack/blas python is using, so i never saw any problems with this combination.

It seems drastic to retroactively change uploaded packages. Is that a normal thing for conda-forge? Was this discussed somewhere?