Naming convention for the different packages on conda

[Luthaf]

Comment:

The metatensor project contains a lot of different sub-packages, developed somehow independently. This feedstock currently builds the native library for metatensor-core, but there are also the following packages:

• native library for metatensor-core
• native library for metatensor-torch (C++ TorchScript interface)
• Python bindings to metatensor-core
• Python bindings to metatensor-torch
• metatensor-operations Python package
• metatensor-learn Python package

We are also leaving open the possibility for more language bindings to metatensor-core or metatensor-torch. The question is what's the best way to split the packages on conda-forge. I think the native libraries should be packaged separately from the Python bindings (they are used on their own when running pre-trained model inside simulation engines).

Our current idea would be to have the following packages on conda-forge:

• metatensor-lib-core => replaces this repository
• metatensor-lib-torch for native metatensor-torch
• metatensor-python-core for Python's metatensor-core
• metatensor-python-torch for Python's metatensor-torch
• metatensor-python-operations
• metatensor-python-learn

As well as the following empty packages with only dependencies on the above (as a convenient way for users to install most packages at the same time):

• metatensor => would give you metatensor-lib-core, metatensor-python-core, metatensor-python-operations and metatensor-python-learn (this would be released with a new, incompatible, version number since we are replacing the package in this repository with a new one)
• metatensor-torch => would give you all of metatensor, plus metatensor-lib-torch, and metatensor-python-torch.

---

For reference, the following packages are currently on PyPI:

• metatensor-core, containing the native library & Python bindings
• metatensor-torch, containing the native metatensor-torch library & Python bindings
• metatensor-operations
• metatensor-learn

As well as an empty metatensor package with dependency on the above. With the proposed package separation above, the following commands would be mostly equivalent:

conda install	pip install
metatensor	metatensor
metatensor-torch	metatensor[torch]
metatensor-python-core	metatensor-core
metatensor-python-torch	metatensor-torch
metatensor-python-operations	metatensor-operations
metatensor-python-learn	metatensor-learn
metatensor-lib-core	/
metatensor-lib-torch	/

---

@conda-forge/core does this looks like a good packaging plan to you? Is the -lib and -python somewhat common in the ecosystem or should we adopt a better naming convention?

[isuruf]

For metatensor-lib-core: libmetatensor-core would be better. For metatensor-python-core: I'd suggest either metatensor-core or python-metatensor-core.

[Luthaf]

Thank you very much for the feedback. The plan will be to go with the following (- package, => dependency):

- libmetatensor

- libmetatensor-torch
  => libmetatensor

- python-metatensor-core
  => libmetatensor

- python-metatensor-torch
  => libmetatensor-torch

- python-metatensor-operation
  => python-metatensor-core

- python-metatensor-learn
  => python-metatensor-core

# Meta packages
- metatensor
  => libmetatensor
  => python-metatensor-core
  => python-metatensor-operations
  => python-metatensor-learn

- metatensor-torch
  => metatensor
  => python-metatensor-torch