Issues

TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 disables PyTorch safe-load default process-wide

ACEsuit/mace#1796 · opened

on Oct 29, 2026

Importing a MACE foundational model messes with Matplotlib global state

ACEsuit/mace#1786 · opened

on Oct 26, 2026

`RealAgnosticResidualNonLinearInteractionBlock` is discontinuous at `r_max` with the default `--apply_cutoff True`

ACEsuit/mace#1784 · opened

on Oct 25, 2026

torch requirement too low for torch.compiler.is_compiling (torch>=1.12 declared, 2.3 needed)

ACEsuit/mace#1782 · opened

on Oct 24, 2026

PolarMACE computes ZBL pair repulsion but omits it from total energy

ACEsuit/mace#1776 · opened

on Oct 21, 2026

DATA-4 — Edge order contract and neighbour-list providers that keep their output on device

ACEsuit/mace#1774 · opened

on Oct 20, 2026

BKD-5 — Interaction-layer span contract and its conformance suite

ACEsuit/mace#1773 · opened

on Oct 20, 2026

huber, universal, l1l2energyforces and dipole/polar losses ignore per-config weights: missing labels trained against zeros, config_weight has no effect

ACEsuit/mace#1767 · opened

on Oct 17, 2026

MACE-POLAR-1-M: energy of a long-wavelength distortion diverges with supercell length (unphysical inter-cell charge transfer)

ACEsuit/mace#1765 · opened

on Oct 16, 2026

Incorrect graph_electrostatics version for MACE-POLAR

ACEsuit/mace#1763 · opened

on Oct 15, 2026

FM-5 — Converter source paths for the dipole, dielectric and magnetic families

ACEsuit/mace#1757 · opened

on Oct 13, 2026

Apple Metal (MPS) kernel backend for MACE: ~10–17× faster inference on Apple silicon — interest, and fp64 conformance on devices without float64

ACEsuit/mace#1754 · opened

on Oct 9, 2026

3

Restarting / resuming training retrains the epoch stored in the checkpoint (off-by-one)

ACEsuit/mace#1747 · opened

on Sep 25, 2026

When restarting/resuming training, the stage two learning rate is loaded from the model checkpoint instead of the command line arguments or configuration file

ACEsuit/mace#1746 · opened

on Sep 25, 2026

1

By default the training procedure will try to evaluate and plot the training data at the end of training

ACEsuit/mace#1745 · opened

on Sep 25, 2026

Loading data from .xyz files can accidentally set forces to zero

ACEsuit/mace#1744 · opened

on Sep 25, 2026

AMSGrad cannot be Disabled

ACEsuit/mace#1742 · opened

on Sep 25, 2026

Multipolar MACELES readouts are missing from optimizer parameter groups

ACEsuit/mace#1741 · opened

on Sep 23, 2026

2

Batches with mixed pbc structures always dispatch to periodic electrostatics

ACEsuit/mace#1735 · opened

on Sep 18, 2026

How is the spin state specified in the .xyz file for MACE-POLAR

ACEsuit/mace#1730 · opened

on Sep 16, 2026

1

Pytorch shape mismatch when running a committee of multiheaded replay fine-tuned foundation models

ACEsuit/mace#1724 · opened

on Sep 11, 2026

Timing MACE with MACE_TIME

ACEsuit/mace#1720 · opened

on Sep 10, 2026

MACE-POLAR-1 (m/l): energies and forces of isolated molecules are not rotationally invariant

ACEsuit/mace#1719 · opened

on Sep 10, 2026

1

create_lammps_model.py fails for mh-1 fine-tuned model with Unknown type name 'o3.Irreps', while mh-0 fine-tuned model works

ACEsuit/mace#1713 · opened

on Sep 8, 2026

ML-IAP inference performs substantially worse than Symmetrix on ARM64 and CUDA13

ACEsuit/mace#1708 · opened

on Sep 1, 2026

3

Navigation Menu

Repository navigation

Issues

TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 disables PyTorch safe-load default process-wide

Importing a MACE foundational model messes with Matplotlib global state

`RealAgnosticResidualNonLinearInteractionBlock` is discontinuous at `r_max` with the default `--apply_cutoff True`

torch requirement too low for torch.compiler.is_compiling (torch>=1.12 declared, 2.3 needed)

PolarMACE computes ZBL pair repulsion but omits it from total energy

DATA-4 — Edge order contract and neighbour-list providers that keep their output on device

BKD-5 — Interaction-layer span contract and its conformance suite

huber, universal, l1l2energyforces and dipole/polar losses ignore per-config weights: missing labels trained against zeros, config_weight has no effect

MACE-POLAR-1-M: energy of a long-wavelength distortion diverges with supercell length (unphysical inter-cell charge transfer)

Incorrect graph_electrostatics version for MACE-POLAR

FM-5 — Converter source paths for the dipole, dielectric and magnetic families

Apple Metal (MPS) kernel backend for MACE: ~10–17× faster inference on Apple silicon — interest, and fp64 conformance on devices without float64

Restarting / resuming training retrains the epoch stored in the checkpoint (off-by-one)

When restarting/resuming training, the stage two learning rate is loaded from the model checkpoint instead of the command line arguments or configuration file

By default the training procedure will try to evaluate and plot the training data at the end of training

Loading data from .xyz files can accidentally set forces to zero

AMSGrad cannot be Disabled

Multipolar MACELES readouts are missing from optimizer parameter groups

Batches with mixed pbc structures always dispatch to periodic electrostatics

How is the spin state specified in the .xyz file for MACE-POLAR

Pytorch shape mismatch when running a committee of multiheaded replay fine-tuned foundation models

Timing MACE with MACE_TIME

MACE-POLAR-1 (m/l): energies and forces of isolated molecules are not rotationally invariant

create_lammps_model.py fails for mh-1 fine-tuned model with Unknown type name 'o3.Irreps', while mh-0 fine-tuned model works

ML-IAP inference performs substantially worse than Symmetrix on ARM64 and CUDA13