feat(pt-expt): add DPA4C-LR model with non-periodic LES/SOG long-rang…

AGENTS.md

@@ -165,6 +165,7 @@ source/               # C++ source code and tests
 - **Configuration**: Use `input_torch.json` format typically
 - **Training**: `dp --pt train input_torch.json`
 - **Requirements**: `torch` package
+- **DPA4C-LR long-range variant**: model type `dpa4c_lr` with fitting type `dpa4c_lr` adds a non-periodic long-range term on top of the compact DPA4C descriptor (`deepmd/pt_expt/model/dpa4c_lr_model.py`, `deepmd/pt_expt/fitting/dpa4c_lr.py`). Two kernels are available: `les` (`erf(alpha*r)/r`, the default) and a trainable sum-of-Gaussians `sog` with multi-channel latent charges. Only non-periodic systems are supported — `box=None` or an all-zero box is accepted, while nonzero periodic cells are rejected — and only the `pt_expt` backend. The long-range all-pairs graph is built outside the compiled region and passed into the graph lower alongside the short-range graph, so `torch.compile` sees no data-dependent neighbor construction; it is built ragged-native with unordered pairs (`i < j`, `build_all_pairs_unordered_graph` in `deepmd/pt_expt/utils/graph_builder.py`) straight from the per-frame node counts — no cutoff search, no padding to the batch max width, and half the directed pair storage (`Σ_f n_f(n_f-1)/2`). `.pt2`/`.pte` freeze is not yet implemented. Tests: `pytest source/tests/pt_expt/model/test_dpa4c_lr.py`.
 
 ### JAX Backend
 

deepmd/dpmodel/descriptor/dpa4c.py

@@ -1452,11 +1452,12 @@ def compute_input_stats(
                 for frame in self._calibration_frames(system):
                     coord = xp.asarray(frame["coord"], dtype=dtype, device=device)
                     atype = xp.asarray(frame["atype"], device=device)
-                    box = (
-                        None
-                        if frame["box"] is None
-                        else xp.asarray(frame["box"], dtype=dtype, device=device)
-                    )
+                    if frame["box"] is None:
+                        box = None
+                    else:
+                        box = xp.asarray(frame["box"], dtype=dtype, device=device)
+                        if xp.all(box == 0):
+                            box = None
                     spin = (
                         None
                         if frame["spin"] is None

deepmd/dpmodel/fitting/__init__.py

@@ -8,6 +8,9 @@
 from .dpa4_ener import (
     SeZMEnergyFittingNet,
 )
+from .dpa4c_lr import (
+    DPA4CLRFitting,
+)
 from .ener_fitting import (
     EnergyFittingNet,
 )
@@ -26,6 +29,7 @@
 
 __all__ = [
     "DOSFittingNet",
+    "DPA4CLRFitting",
     "DipoleFitting",
     "EnergyFittingNet",
     "InvarFitting",