Metatensor models have import error due to 2.7 PyTorch but we get an uninformative error message.

[CompRhys]

The current solution to top level imports of models doesn't give useful error messages.

Lazy imports were lumped into #179 but that also doesn't seem to change anything. Unsure haven't checked docs build with the lazy import route.

[CompRhys]

ImportError: Trying to load metatensor-torch with torch v2.7.0+cu126, we found builds for torch v2.3, v2.2, v2.0, v2.6, v1.12, v2.4, v1.13, v2.5, v2.1; which are not ABI compatible.

[Luthaf]

The latest release of metatensor_torch (v0.7.6) is now compatible with torch 2.7. As mentioned in metatensor/metatensor#913, we are pinning the torch requirement for our wheels. Maybe the failure comes from a bug/a different behavior between pip and uv?

EDIT: the CI logs would be more useful if uv run --with . would print the exact version of packages it installs
EDIT2: metatrain is still pinning metatensor-torch v0.7.5, we might want to relax this a bit @frostedoyster

[Luthaf]

This seems to be an issue with how uv run --with does dependency resolution: astral-sh/uv#13173

[janosh]

@Luthaf thanks for troubleshooting and reporting this upstream! 🚀

[CompRhys]

Reopening this because we're still seeing errors and it was closed by linked PR not manually.

[Luthaf]

Might be that the code/example needs an upgrade to the latest metatrain (v2025.6)