fix(lammps): support 2025 Kokkos API (deepmodeling#5987)

## Summary

Backport `0002-support-lammps-2025-kokkos-api.patch` from
`conda-forge/deepmd-kit-feedstock@d02138263b0b30629fe28c68c8bcecc2b8644513`.

Support both the LAMMPS 22Jul2025 Kokkos API and newer LAMMPS releases
by selecting the appropriate reverse-communication buffer and per-atom
energy accumulator types from `LAMMPS_VERSION_NUMBER`.

## Validation

- Four-file backport matching the feedstock patch.
- Split from deepmodeling#5985 so the compatibility change can be reviewed
independently.
- Full build is left to upstream CI.

Agent: ChatGPT
Model: GPT-5.6 Sol

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

- **Compatibility**
- Improved support across multiple LAMMPS versions for Kokkos-based
DeepMD and DPA-4-spin simulations.
- Updated reverse communication and per-atom energy handling to use
version-compatible data formats.

- **Bug Fixes**
- Resolved compatibility issues that could affect reverse communication
and energy accumulation across LAMMPS releases.
- Preserved existing force, magnetic-force, and virial calculation
behavior.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Author: njzjz-bot

source/lmp/pair_deepmd_kokkos.cpp

@@ -113,7 +113,7 @@ void PairDeepMDKokkos<DeviceType>::unpack_reverse_comm(int n,
 
 template <class DeviceType>
 int PairDeepMDKokkos<DeviceType>::pack_reverse_comm_kokkos(
-    int n, int first, DAT::tdual_double_1d& buf) {
+    int n, int first, DeepMDKokkosCommBuffer& buf) {
   auto d_buf = buf.template view<DeviceType>();
   if (reverse_virial) {
     auto reverse_virial_data = k_reverse_virial.template view<DeviceType>();
@@ -141,7 +141,7 @@ int PairDeepMDKokkos<DeviceType>::pack_reverse_comm_kokkos(
 
 template <class DeviceType>
 void PairDeepMDKokkos<DeviceType>::unpack_reverse_comm_kokkos(
-    int n, DAT::tdual_int_1d list, DAT::tdual_double_1d& buf) {
+    int n, DAT::tdual_int_1d list, DeepMDKokkosCommBuffer& buf) {
   auto d_buf = buf.template view<DeviceType>();
   auto d_list = list.template view<DeviceType>();
   if (reverse_virial) {

source/lmp/pair_deepmd_kokkos.h

@@ -28,6 +28,14 @@ PairStyle(deepmd/kk/host,PairDeepMDKokkos<LMPHostType>);
 
 namespace LAMMPS_NS {
 
+// LAMMPS 22Jul2025 exposes reverse-communication buffers as X_FLOAT; starting
+// with 10Sep2025, Kokkos pair styles use a fixed double buffer.
+#if LAMMPS_VERSION_NUMBER < 20250910
+using DeepMDKokkosCommBuffer = DAT::tdual_xfloat_1d;
+#else
+using DeepMDKokkosCommBuffer = DAT::tdual_double_1d;
+#endif
+
 // GPU-resident inference for exported ``.pt2`` models whose forward consumes
 // an explicit edge graph: both the graph-input form (a compact, unpadded
 // neighbor graph) and the edge-input form. Both are dispatched through
@@ -64,10 +72,10 @@ class PairDeepMDKokkos : public PairDeepMD, public KokkosBase {
   // host-staged path.
   int pack_reverse_comm(int, int, double*) override;
   void unpack_reverse_comm(int, int*, double*) override;
-  int pack_reverse_comm_kokkos(int, int, DAT::tdual_double_1d&) override;
+  int pack_reverse_comm_kokkos(int, int, DeepMDKokkosCommBuffer&) override;
   void unpack_reverse_comm_kokkos(int,
                                   DAT::tdual_int_1d,
-                                  DAT::tdual_double_1d&) override;
+                                  DeepMDKokkosCommBuffer&) override;
 
   // Build the device edge graph of the edge-input schema from the Kokkos full
   // neighbor list, returning the edge count. Public because it launches
@@ -101,8 +109,14 @@ class PairDeepMDKokkos : public PairDeepMD, public KokkosBase {
 
   // Per-atom energy accumulator (aliases the base Pair ``eatom`` host array so
   // downstream per-atom computes/dumps see it after the device-to-host sync).
+  // The transformed accumulator view was added in the 10Sep2025 release.
+#if LAMMPS_VERSION_NUMBER < 20250910
+  DAT::tdual_double_1d k_eatom;
+  typename AT::t_double_1d d_eatom;
+#else
   DAT::ttransform_kkacc_1d k_eatom;
   typename AT::t_kkacc_1d d_eatom;
+#endif
 
   int edge_capacity;       // allocated edges in d_edge_index / d_edge_vec
   bool edge_vec_fp32;      // model graph ABI consumes edge vectors in fp32

source/lmp/pair_dpa4spin_kokkos.cpp

@@ -115,7 +115,7 @@ void PairDPA4SpinKokkos<DeviceType>::unpack_reverse_comm(int n,
 
 template <class DeviceType>
 int PairDPA4SpinKokkos<DeviceType>::pack_reverse_comm_kokkos(
-    int n, int first, DAT::tdual_double_1d& buf) {
+    int n, int first, DPA4SpinKokkosCommBuffer& buf) {
   auto d_buf = buf.template view<DeviceType>();
   if (reverse_virial) {
     auto reverse_virial_data = k_reverse_virial.template view<DeviceType>();
@@ -147,7 +147,7 @@ int PairDPA4SpinKokkos<DeviceType>::pack_reverse_comm_kokkos(
 
 template <class DeviceType>
 void PairDPA4SpinKokkos<DeviceType>::unpack_reverse_comm_kokkos(
-    int n, DAT::tdual_int_1d list, DAT::tdual_double_1d& buf) {
+    int n, DAT::tdual_int_1d list, DPA4SpinKokkosCommBuffer& buf) {
   auto d_buf = buf.template view<DeviceType>();
   auto d_list = list.template view<DeviceType>();
   if (reverse_virial) {

source/lmp/pair_dpa4spin_kokkos.h

@@ -28,6 +28,14 @@ PairStyle(dpa4spin/kk/host,PairDPA4SpinKokkos<LMPHostType>);
 
 namespace LAMMPS_NS {
 
+// LAMMPS 22Jul2025 exposes reverse-communication buffers as X_FLOAT; starting
+// with 10Sep2025, Kokkos pair styles use a fixed double buffer.
+#if LAMMPS_VERSION_NUMBER < 20250910
+using DPA4SpinKokkosCommBuffer = DAT::tdual_xfloat_1d;
+#else
+using DPA4SpinKokkosCommBuffer = DAT::tdual_double_1d;
+#endif
+
 // GPU-resident inference for exported native-spin ``.pt2`` models whose forward
 // consumes the compact canonical graph: a dual-CSR neighbor topology with
 // uint32 indices and float32 edge vectors, plus the per-node magnetic moment.
@@ -64,10 +72,10 @@ class PairDPA4SpinKokkos : public PairDPA4Spin, public KokkosBase {
   // overrides serve the host-staged path.
   int pack_reverse_comm(int, int, double*) override;
   void unpack_reverse_comm(int, int*, double*) override;
-  int pack_reverse_comm_kokkos(int, int, DAT::tdual_double_1d&) override;
+  int pack_reverse_comm_kokkos(int, int, DPA4SpinKokkosCommBuffer&) override;
   void unpack_reverse_comm_kokkos(int,
                                   DAT::tdual_int_1d,
-                                  DAT::tdual_double_1d&) override;
+                                  DPA4SpinKokkosCommBuffer&) override;
 
   // Gather the per-node magnetic moment from the Kokkos ``sp`` array. Public
   // because it launches an extended device lambda, which CUDA forbids inside
@@ -94,8 +102,14 @@ class PairDPA4SpinKokkos : public PairDPA4Spin, public KokkosBase {
 
   // Per-atom energy accumulator (aliases the base Pair ``eatom`` host array so
   // downstream per-atom computes/dumps see it after the device-to-host sync).
+  // The transformed accumulator view was added in the 10Sep2025 release.
+#if LAMMPS_VERSION_NUMBER < 20250910
+  DAT::tdual_double_1d k_eatom;
+  typename AT::t_double_1d d_eatom;
+#else
   DAT::ttransform_kkacc_1d k_eatom;
   typename AT::t_kkacc_1d d_eatom;
+#endif
 
   bool reverse_virial;     // reverse communication operates on centroid virial
   bool reverse_used_host;  // force reverse communication selected host staging