OpenEquivariance ML-IAP compilation

[hmcezar]

Hi,

I'm trying to compile an Allegro model I have trained with ML-IAP compilation and OpenEquivariance.
The compilation without the --modifier works, but when I try to use --modifiers enable_OpenEquivariance I get:

Traceback (most recent call last):
  File "/opt/lammps/bin/nequip-prepare-lmp-mliap", line 8, in <module>
    sys.exit(main())
             ^^^^^^
  File "/opt/lammps/lib/python3.11/site-packages/nequip/integrations/lammps_mliap/create_lmp_mliap_file.py", line 79, in main
    mliap_module = NequIPLAMMPSMLIAPWrapper(
                   ^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/lammps/lib/python3.11/site-packages/nequip/integrations/lammps_mliap/lmp_mliap_wrapper.py", line 80, in __init__
    raise ValueError(
ValueError: Provided modifier `enable_OpenEquivariance` is not available in the model; only the following are available for the provided model: ['enable_CuEquivarianceContracter', 'enable_TritonContracter', 'modify_PerTypeScaleShift']

I didn't train the model using OpenEquivariance, not sure if this matters.
I am trying to compile on AMD MI250x GPUs. I'm using the patch_10Sep2025 version of LAMMPS, and this is my environment:

nequip                    0.15.0
nequip-allegro            0.7.1
openequivariance          0.4.1
torch                     2.7.0+rocm6.2.4

Am I missing something?

Thanks!

[kavanase]

Hi @hmcezar,
OpenEquivariance acceleration is not possible with Allegro models, for training or inference. CuEquivariance or Triton accelerations are possible for the tensor products in Allegro, see https://nequip.readthedocs.io/projects/allegro/en/latest/guide/accelerations.html