[kavanase]

_from_saved_model, used by NequIPCalculator and NequIPTorchSimCalc to load checkpoints and packages in eager mode, puts the model in eval mode but leaves its parameters with requires_grad=True. Other inference paths (TorchScript, AOTInductor and LAMMPS ML-IAP) already freeze them with prepare_model_for_compile. With trainable parameters, autograd keeps the tensors needed for parameter gradients while computing forces, although inference never computes those gradients. This PR adds model.requires_grad_(False) after model.eval() to instead freeze the parameters, which significantly reduces memory demand.

Peak RSS in eager CPU inference on a rattled 216-atom NaCl cell; Apple M5 Pro, 6 threads, torch 2.10:

Model	before	after
Allegro-OAM-M	3.65 GiB	2.99 GiB
Allegro-OAM-M, with einsum allegro	2.37 GiB	1.76 GiB
Allegro-OAM-L, with einsum allegro	10.22 GiB	6.12 GiB
NequIP-OAM-S	0.52 GiB	0.52 GiB
NequIP-OAM-M	1.06 GiB	0.98 GiB

'einsum allegro' is the Contracter einsum change in mir-group/allegro#148. For Allegro, the saving is one [edges, channels, i, j] tensor per layer, kept for the Contracter weight gradient: 256 KiB per edge for Allegro-OAM-L, about 4 GiB here.

NequIP models save less, since their large per-edge tensors are needed for the forces anyway, and only the radial MLP's hidden activations and some node-level tensors are dropped: peak mem is 3% (NequIP-OAM-M) and 5% (NequIP-OAM-L) lower on 216 atoms, and 5.5% and 6.5% lower on a 1,000-atom cell:

Model	before	after
NequIP-OAM-M	2.55 GiB	2.41 GiB (−5.5%)
NequIP-OAM-L	6.19 GiB	5.79 GiB (−6.5%)

Tested:

• Energy, forces, stress and second derivatives are identical frozen vs unfrozen for Allegro-OAM-S, NequIP-OAM-S and freshly built models.

Types of changes

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds or improves functionality)
• Breaking change (fix or feature that would cause existing functionality to change)
• Documentation improvement (updates to user guides, docstrings, or developer docs)

Checklist:

• I have read "Contributing to NequIP".
• My code follows the code style of this project and has been formatted using black.
• All new and existing tests passed, including on GPU (if relevant).
• I have added tests that cover my changes (if relevant).
• I have updated CHANGELOG.md.
• I have updated the documentation (if relevant).

[kavanase]

(Pre-existing test failure: https://github.com/mir-group/nequip/actions/runs/39305994111/job/118243278242)