fix: repair stale post-DGL-removal refs in LAMMPS export + CI

The LAMMPS export wrapper and its test predated the DGL-backend removal
and module renames, breaking the Lint and Test CI jobs:

- _lammps.py imported the removed `_pyg`-suffixed modules
  (graph._compute_pyg, apps._pes_pyg, models._m3gnet_pyg/_tensornet_pyg);
  repoint to graph._compute, apps.pes, models._m3gnet/_tensornet.
- test_lammps_export.py guarded on the removed `matgl.config.BACKEND`
  and imported the same dead modules; drop the guard and import from the
  public APIs.
- cli.py: cover the `attr-defined` code in the create-lammps-model
  type-ignore so mypy passes.

Also correct the READMEs: M3GNet (not just TensorNet) is supported by
the LAMMPS export / Kokkos path, and drop the outdated AdvancedSoft
third-party reference.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: shyuep

README.md

@@ -118,11 +118,11 @@ pip install matgl[jax]    # JAX/XLA inference for TensorNet/QET (CPU / CUDA / Ap
 
 ## LAMMPS support
 
-MatGL ships a native LAMMPS `pair_style matgl` that evaluates a TorchScript-compiled MatGL **TensorNet** PES
-(energies, forces, and virials) on every timestep via LibTorch, with both a CPU/serial pair style and a Kokkos
-GPU/host variant (`pair_style matgl/kk`). Export a LAMMPS-loadable model from any pre-trained TensorNet with
-`mgl create-lammps-model`. Build and usage instructions, the input-deck syntax, and current limitations are in
-[`lammps/README.md`](lammps/README.md).
+MatGL ships a native LAMMPS `pair_style matgl` that evaluates a TorchScript-compiled MatGL **TensorNet** or
+**M3GNet** PES (energies, forces, and virials) on every timestep via LibTorch, with both a CPU/serial pair style
+and a Kokkos GPU/host variant (`pair_style matgl/kk`). Export a LAMMPS-loadable model from a pre-trained TensorNet
+or M3GNet with `mgl create-lammps-model`. Build and usage instructions, the input-deck syntax, and current
+limitations are in [`lammps/README.md`](lammps/README.md).
 
 ## Docker images
 

lammps/README.md

@@ -1,7 +1,7 @@
 # MatGL → LAMMPS pair_style
 
 `pair_matgl` is a LAMMPS pair style that loads a TorchScript-compiled
-**MatGL TensorNet** PES (PyG backend, no-Warp, extensive head) and uses
+**MatGL TensorNet or M3GNet** PES (PyG backend, extensive head) and uses
 LibTorch to evaluate energies, forces, and the virial tensor on every
 timestep.
 
@@ -144,9 +144,14 @@ Currently a no-op.
   proportionally to the number of ranks.
 - **No restart support.** The model lives on disk; `restart` files don't
   capture the path. Re-issue `pair_style` / `pair_coeff` after a restart.
-- **TensorNet only.** The pair style loads a TorchScript-compiled
-  TensorNet PES. Other architectures (M3GNet, CHGNet, MEGNet, SO3Net,
-  QET) are not yet wired into the LAMMPS export path.
+- **TensorNet and M3GNet only.** The export path (`mgl
+  create-lammps-model` / `LAMMPSMatGLModel`) supports the PyG TensorNet
+  and M3GNet PES models with an extensive head (TensorNet must be
+  no-Warp; both require `use_smooth=True`, and M3GNet requires
+  `use_phi=False`). For M3GNet the three-body line graph is built inside
+  the TorchScript module, so both the CPU and Kokkos pair styles run it
+  with no extra handling. CHGNet, MEGNet, SO3Net, and QET are not yet
+  wired into the LAMMPS export path.
 
 ## Continuous integration
 

src/matgl/cli.py

@@ -241,7 +241,7 @@ def create_lammps_model(args: argparse.Namespace) -> int:
     print("  r_max     :", wrapper.r_max)
     print("  n_species :", wrapper.n_species)
     print("  dtype     :", args.dtype)
-    species = list(potential.model.element_types)  # type:ignore[union-attr,arg-type]
+    species = list(potential.model.element_types)  # type:ignore[union-attr,arg-type,attr-defined]
     print("  species   :", species[: wrapper.n_species])
     return 0
 

src/matgl/ext/_lammps.py

@@ -24,7 +24,7 @@
 from torch import Tensor, nn
 from torch.autograd import grad
 
-from matgl.graph._compute_pyg import (
+from matgl.graph._compute import (
     compute_pair_vector_and_distance,
     compute_theta_and_phi,
     create_line_graph_torch,
@@ -34,7 +34,7 @@
 from matgl.utils.maths import decompose_tensor, tensor_norm
 
 if TYPE_CHECKING:
-    from matgl.apps._pes_pyg import Potential
+    from matgl.apps.pes import Potential
 
 logger = logging.getLogger(__name__)
 
@@ -343,8 +343,8 @@ def __init__(
 
         # Imports kept local so the public ``matgl.ext`` namespace doesn't
         # hard-require these submodules at import time.
-        from matgl.models._m3gnet_pyg import M3GNet
-        from matgl.models._tensornet_pyg import TensorNet
+        from matgl.models._m3gnet import M3GNet
+        from matgl.models._tensornet import TensorNet
 
         model = potential.model
 

tests/ext/test_lammps_export.py

@@ -11,20 +11,14 @@
 import numpy as np
 import pytest
 import torch
-
-import matgl
-
-if matgl.config.BACKEND != "PYG":
-    pytest.skip("LAMMPS export only supports PyG backend", allow_module_level=True)
-
 from pymatgen.core import Lattice, Structure
 from pymatgen.optimization.neighbors import find_points_in_spheres
 
-from matgl.apps._pes_pyg import Potential
-from matgl.ext._lammps import LAMMPSMatGLModel
-from matgl.ext._pymatgen_pyg import Structure2Graph
-from matgl.models._m3gnet_pyg import M3GNet
-from matgl.models._tensornet_pyg import TensorNet
+import matgl
+from matgl.apps.pes import Potential
+from matgl.ext.lammps import LAMMPSMatGLModel
+from matgl.ext.pymatgen import Structure2Graph
+from matgl.models import M3GNet, TensorNet
 
 
 def _build_lammps_inputs(structure: Structure, element_types: tuple[str, ...], cutoff: float, dtype: torch.dtype):