[Bug] esolver_type=dp (MDCell): message-passing models (DPA-2/repformers) abort at MD step 1 or give wrong energies — missing InputNlist communication setup #8108

Summary

esolver_type = dp with a message-passing model (DPA-2 / repformers) fails on the MD (MDCell) path in v3.11.0-beta10:

• with MPI enabled (the normal build), the run aborts at MD step 1:
  MPI_Comm_rank on an invalid communicator (MPI_ERR_COMM), and
• with the abort worked around (no-op MPI interposer / non-MPI setup), the run completes but the energy and forces deviate massively from the model's own prediction (here: energy off by ~ +36.6 eV on a 50-atom cell, forces off by ~0.5 eV/Å rms). se_e2_a models are unaffected.

Root cause

The MDCell branch of ESolver_DP::runner() constructs deepmd::InputNlist without any communication metadata (nswap = 0, no world). Message-passing descriptors (repformers) call torch.ops.deepmd.border_op(...) in every descriptor layer whenever a comm_dict is present, and DeepPotPT::compute builds that comm_dict from the InputNlist (lmp_list.world, sendlist, sendnum, sendproc, recvproc, nswap). Consequences:

1. comm_dict["communicator"] is 0 -> border_op calls MPI_Comm_rank(NULL) when MPI is initialized -> abort.
2. Without MPI the op silently no-ops -> the forward ghost-embedding exchange never happens; the message-passing layers aggregate zero embeddings for ghost neighbors -> wrong energies/forces.

Related detail found while debugging: the reverse force accumulation is not the esolver's job — MD_func::force_virial() already calls DomainDecomposition::accumulate_ghost_forces() after runner() returns. The esolver must return the raw model output (partial forces on owned atoms + the model's ghost force entries); doing accumulation inside the esolver double-counts.

For reference, se_e2_a works because it has no message passing, and cells with nghost == 0 never hit this path. This is independent of the neighbor-ordering bug fixed in #8071 / #8104.

Proposed fix (single-rank)

In source/source/source_esolver/esolver_dp.cpp (runner(), MDCell branch), build the ghost -> owned source mapping (atom.type, atom.type_index) -> owned index and pass the comm-aware InputNlist constructor:

nswap = (nghost > 0) ? 1 : 0
sendnum = recvnum = {nghost}
sendproc = recvproc = {0}          // self-exchange
sendlist = {ghost -> owned sources}
world = &dp_world_comm             // address of an MPI_Comm variable

i.e. the same self-exchange topology the LAMMPS deepmd pair style provides in the single-rank case (world must point to an MPI_Comm variable — passing the value itself crashes inside build_comm_dict).

This implements the single-rank case only; multi-rank DP-MD would additionally need a real cross-rank exchange topology (cf. DomainDecomposition::ghost_slots_).

Verification (with the fix)

Same job, 10 MD steps, every frame of OUT.ABACUS/MD_dump vs deepmd.infer.DeepPot.eval:

frame 0..9: rmse = 2e-05 eV/A,  max|dF| <= 8e-05 eV/A,  energy within ~0.01 eV

Reproduction package

https://gist.github.com/MarkYoungIV/b692ca7288ebbef0d332f271b4414445 (parts because gists don't take binaries)

cat part_aa part_ab part_ac > dpa2_repro.zip.b64
base64 -d dpa2_repro.zip.b64 > dpa2_repro.zip
unzip dpa2_repro.zip

Contains: INPUT (10-step MD, esolver_type dp), STRU (50-atom LiGePS cell), frozen_sse_abacus.pth (a DPA-2/repformers single-task model with public weights, built from the public DPA-2.4-7M release — SSE_ABACUS branch), esolver_dp_comm_fix.patch, and verify_after_fix.py (frame-by-frame comparison). On a pristine beta10 build the job aborts at MD step 1 as described; after the patch it completes and matches the Python API to ~2e-05 eV/Å.

Related: #8071 (neighbor ordering, fixed by #8104 — the patch here does not overlap that hunk).