pair_style deepmd/kk does not populate global virial / thermo pressure for a DPA4 pt_expt model

[BDxka]

Bug summary

When using a converted DPA4 pt_expt model with LAMMPS, pair_style deepmd/kk produces correct per-atom virials but fails to populate the global virial used by LAMMPS thermo pressure. As a result, Press, Pxx, Pyy, and Pzz contain only the kinetic contribution; at zero velocity they are exactly zero.
The same structure and model work correctly with the non-Kokkos path, pair_style deepmd.
This is important for NPT simulations: the barostat receives an incorrect pressure signal, which can cause unphysical cell evolution.

DeePMD-kit Version

Source checkout at commit 14a71f1

Backend and its version

PyTorch 2.11.0+cu126 (CUDA 12.6)

How did you download the software?

Built from source

Input Files, Running Commands, Error Log, etc.

Environment

• LAMMPS: patch_4Jul2026 (LAMMPS (4 Jul 2026))

• GPU: NVIDIA A40

• CUDA: 12.6

• DeePMD-kit was built from source at commit 14a71f13.

• The DPA4 model was converted with:

  dp convert-backend --atomic-virial DPA4-Pro-MPtrj.pt DPA4-Pro-converted.pt2

Minimal test

The same relaxed alpha-quartz SiO2 structure was evaluated with zero velocities and run 0.
Input:

read_data quartz_dpa_relaxed.lmp
mass 1 28.0855
mass 2 15.9994
pair_style deepmd/kk DPA4-Pro-converted.pt2
pair_coeff * * Si O
compute stress all centroid/stress/atom NULL virial
thermo_style custom step temp press pxx pyy pzz
run 0

Observed behavior

With pair_style deepmd/kk, thermo output is exactly zero at zero velocity:
Press = 0
Pxx = 0
Pyy = 0
Pzz = 0

However, compute centroid/stress/atom returns nonzero per-atom virials. Their sum agrees with an independent ASE stress calculation within approximately 0.1%.
For the same structure and converted model, replacing only:
pair_style deepmd/kk DPA4-Pro-converted.pt2
with:
pair_style deepmd DPA4-Pro-converted.pt2
produces nonzero thermo pressure/stress consistent with both ASE and the summed per-atom virial.
Therefore, the issue appears specific to the Kokkos path.

Steps to Reproduce

1. Convert a DPA4-Pro-MPtrj checkpoint to pt_expt format with:
   dp convert-backend --atomic-virial DPA4-Pro-MPtrj.pt DPA4-Pro-converted.pt2
2. Build LAMMPS with Kokkos/CUDA support against this DeePMD-kit checkout.
3. Run the input below (relaxed alpha-quartz SiO2, 72 atoms) with
   pair_style deepmd/kk, using:
   lmp -k on g 1 -sf kk -pk kokkos newton on neigh half -in in.lammps
4. Observe that thermo Press/Pxx/Pyy/Pzz are exactly zero, while
   compute centroid/stress/atom still gives correct nonzero per-atom
   virials (summed, they agree with an independent ASE calculation
   within ~0.1%).
5. Repeat step 3 with pair_style deepmd (no /kk) instead, changing
   nothing else. Thermo pressure/stress becomes correct and nonzero.

Further Information, Files, and Links

Source-level observation:

• source/lmp/pair_deepmd.cpp performs the explicit global-virial reduction under if (vflag).
• source/lmp/pair_deepmd_kokkos.cpp performs the corresponding reduction under if (vflag_global).
  In the failing reproduction, vflag is nonzero (10 when the per-atom virial compute is requested), whereas vflag_global = 0. Therefore, the explicit Kokkos reduction is skipped. The per-atom virial buffer is nevertheless nonzero and agrees with the non-Kokkos result; the LAMMPS global virial[] / thermo pressure remains zero.
  I have attached a zip archive containing the minimal structure, the two LAMMPS input files, the two comparison logs, and a README. The converted model is not included because it is large; it can be regenerated from the DPA4-Pro-MPtrj checkpoint using the conversion command given above.
  The DEBUG_VFLAG lines in the attached logs were added locally for diagnosis and are not part of the upstream DeePMD-kit output.

deepmd_kk_virial_bug_report.zip

[dosubot]

Hi @BDxka! I'm Dosu and I'm helping the DeepModeling team.

Your diagnosis is spot-on. The root cause is a guard mismatch between the two pair styles:

• pair_deepmd.cpp accumulates the global virial under if (vflag) — the raw bitmask that is nonzero whenever any virial (global or per-atom) is requested [1].
• pair_deepmd_kokkos.cpp guards the same reduction with if (vflag_global), which is only set when the global virial is explicitly requested [2].

Both paths call ev_init(eflag, vflag), which derives vflag_global, vflag_atom, and cvflag_atom from the raw vflag bitmask. In your scenario, adding compute centroid/stress/atom sets cvflag_atom = 1 and makes vflag nonzero, but vflag_global can remain 0 depending on how LAMMPS's implicit pressure compute propagates back through the Kokkos pair path. The non-Kokkos path works because if (vflag) catches this case; the Kokkos path's narrower if (vflag_global) does not, so the virial[] array stays zeroed and thermo pressure reports only the kinetic contribution.

The pair_deepmd_kokkos.cpp file was introduced in PR #5758, so this code path is quite new and this edge case likely wasn't caught.

Suggested fix: change the Kokkos guard to match the non-Kokkos behaviour:

// Before
if (vflag_global) {

// After
if (vflag_either) {   // or simply: if (vflag)

vflag_either is the idiomatic LAMMPS choice (true when either global or per-atom virial is needed). The extra work of reducing the global virial when only per-atom is requested is negligible next to model inference cost.

Would you be able to test that one-line change locally and confirm it resolves the issue? That would strengthen a fix PR.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.

[BDxka]

I tested the suggested change locally.

I changed the guard in source/lmp/pair_deepmd_kokkos.cpp from:
if (vflag_global)
to:
if (vflag_either)

After an incremental rebuild and reinstall of LAMMPS, the same pair_style deepmd/kk reproduction case now reports nonzero global thermo stress:
Press = 159.18387 bar
Pxx = 232.23730 bar
Pyy = 231.30624 bar
Pzz = 14.00809 bar

The summed per-atom centroid virial remains unchanged and agrees with the previous non-Kokkos result. This resolves the issue in my test case. Thank you.