[BUG] deepmd/kk silently ignores spin atoms instead of rejecting them

[njzjz-bot]

Summary

The host deepmd pair style rejects spin atoms and tells the user to select a spin-aware pair style. The deepmd/kk compute override omits that guard, so an ordinary model can run with atom_style spin/kk while silently ignoring every atom's spin state.

This is present on origin/master at 8cfd46e37448.

Evidence

PairDeepMD::compute() contains:

if (atom->sp_flag) {
  error->all(FLERR,
             "Pair style 'deepmd' does not support spin atoms, please use "
             "pair style 'deepspin' instead.");
}

PairDeepMDKokkos::compute() overrides the entire host compute path and never performs this check. Its device graph is built only from coordinates and atom types; it neither consumes atom->sp nor produces atom->fm.

Trigger and impact

Run an ordinary edge-, graph-, or canonical-input .pt2 model with:

atom_style spin/kk
pair_style deepmd/kk <ordinary-model.pt2>

Unlike pair_style deepmd, initialization and compute proceed. Changing magnetic moments has no effect, and no magnetic force is produced, which can silently yield invalid spin dynamics.

Expected behavior

PairDeepMDKokkos::init_style() or compute() should apply the same atom->sp_flag rejection as the host path. A regression should compare deepmd and deepmd/kk behavior for a spin atom style.

---

Coding agent: Codex
Codex version: codex-cli 0.149.0
Model: gpt-5.6-sol
Reasoning effort: xhigh

[dosubot]

I confirmed the bug by reading pair_deepmd_kokkos.cpp on master directly.

The PairDeepMDKokkos::compute() method (web_page:5494d20d) begins with:

ev_init(eflag, vflag);

if (vflag_atom) {
  error->all(FLERR, "6-element atomic virial is not supported ...");
}

There is no atom->sp_flag check anywhere in compute() or init_style(). By contrast, PairDeepMD::compute() checks at the very top :

if (atom->sp_flag) {
  error->all(FLERR,
             "Pair style 'deepmd' does not support spin atoms, please use "
             "pair style 'deepspin' instead.");
}

The same guard also exists in pack_reverse_comm() and unpack_reverse_comm() of the host path, but PairDeepMDKokkos overrides both methods with its own Kokkos device/virial logic and neither copy includes the sp_flag check [1].

The fix is straightforward — add the guard near the top of PairDeepMDKokkos::compute(), right after ev_init:

if (atom->sp_flag) {
  error->all(FLERR,
             "Pair style 'deepmd/kk' does not support spin atoms, please "
             "use pair style 'deepspin' instead.");
}

Optionally the same check can be added to init_style() so it fails at setup time rather than at the first step. Note that init_style() calls PairDeepMD::init_style() which does not contain the spin guard either — that guard only lives in compute() — so adding it to init_style() would be a net improvement for both paths.

This is one of several parity gaps between deepmd and deepmd/kk that have surfaced recently (e.g. the vflag_global vs vflag_either virial issue in #5983 and the GPU device mismatch in #5994), all traceable to the Kokkos path introduced in PR #5758.

To reply, just mention @dosu.

---

Share context across your team and agents. Try Dosu.