v0.8.2

[0.8.2]

Added

• per-type average number of neighbors normalization option by @gharari1

Changed

• minor changes to TritonContracter
• minimum nequip version is v0.17.0

v0.8.1

[0.8.1]

Fixed

• #138

v0.8.0

[0.8.0]

Removed

• [Breaking] AllegroEnergyModel has been removed. Energy-only models can be constructed by using AllegroModel with do_derivatives=False.

Fixed

• #134

v0.7.1

[0.7.1]

Fixed

• nequip-package will now always pick up Triton (even if packaging without GPUs) and CuEq files (even if CuEq is not installed at package-time)

Changed

• bound nequip version >=0.13.0

v0.7.0: CuEquivariance Support

We are proud to announce the integration of NVIDIA's CuEquivariance to accelerate Allegro models. Information on how to use it can be found in the CuEquivariance Acceleration page.

Training

CuEquivariance can be used to accelerate training by using nequip.model.modify in the config file:

training_module:
  _target_: allegro.train.EMALightningModule
  # ... other training module configurations ...
  model:
    _target_: nequip.model.modify
    modifiers:
      - modifier: enable_CuEquivarianceContracter
    model:
      _target_: allegro.model.AllegroModel
      # ... your standard Allegro model configuration ...
      compile_mode: compile
      # ^ CuEquivariance composes with train-time compilation

Inference

CuEquivariance can be used to accelerate TorchScript and AOTInductor inference in ASE by doing nequip-compile with --modifiers enable_CuEquivarianceContracter.

AOT Inductor Compilation

nequip-compile \
    path/to/model.ckpt \
    path/to/compiled_model.nequip.pt2 \
    --device cuda \
    --mode aotinductor \
    --target ase \
    --modifiers enable_CuEquivarianceContracter

TorchScript Compilation

nequip-compile \
    path/to/model.ckpt \
    path/to/compiled_model.nequip.pth \
    --device cuda \
    --mode torchscript \
    --target ase \
    --modifiers enable_CuEquivarianceContracter

v0.6.3

[0.6.3]

Fixed

• enable_TritonContracter restores weights correctly now

v0.6.2

[0.6.2]

Added

• Exposed AllegroEnergyModel in docs

v0.6.1

[0.6.1]

Removed unused code and revamped docs.

Removed

• scatter_features option (undocumented feature)

v0.6.0

[0.6.0]

Changed

• bound nequip version >=0.8.0
• [Breaking] mechanism for using custom Triton TP kernel

v0.5.1

[0.5.1]

Changed

• bound nequip version under 0.8.0