This issue comes from a Codex global scan of deepmodeling/deepmd-kit at commit 73de44b1f94471b2e3bdb6b11f57b34d7bc791bb.

Problem

dprc_pairwise_map_cpu() reports every local atom as a local QM atom when the pairwise DPRc index map contains only one non-placeholder fragment.

The mapper groups atoms by fragment index, ignoring -1 placeholders:

• deepmd-kit/source/lib/src/pairwise.cc

  Lines 21 to 40 in 73de44b

  	void deepmd::group_atoms_cpu(std::vector<std::vector<int>>& fragments,
  	const std::vector<int>& idxs) {
  	int natoms = idxs.size();
  	// sort idxs
  	std::vector<size_t> idxs_idx = sort_indexes(idxs);
  	// now idxs_idx is sorted index, so we can easily group atoms in only one loop
  	int last_frag = -1;
  	for (size_t ii = 0; ii < idxs.size(); ii++) {
  	int frag = idxs[idxs_idx[ii]];
  	if (frag == -1) {
  	// -1 is the place holder
  	continue;
  	}
  	if (frag != last_frag) {
  	last_frag = frag;
  	fragments.emplace_back();
  	}
  	// push to the last fragment
  	fragments.back().push_back(idxs_idx[ii]);
  	}

The pairwise map then initializes nqm_real to nloc for the nfragments == 1 case:

• deepmd-kit/source/lib/src/pairwise.cc

  Lines 91 to 99 in 73de44b

  	int max_fragment_size = max_fragment_real_size + max_fragment_ghost_size;
  	int map_size = nqm + max_fragment_real_size + max_fragment_ghost_size;
  	// (3, 4, 0, 1, 2, 10, 11),
  	// (3, 4, 5, 6, 7, 10, -1),
  	// (3, 4, 8, 9, -1, 10, -1)
  	forward_qmmm_map.resize(static_cast<size_t>(nfragments - 1) * map_size);
  	std::fill(forward_qmmm_map.begin(), forward_qmmm_map.end(), -1);
  	int nqm_real = nloc; // init for nfragments = 1
  	for (int ii = 0; ii < nfragments - 1; ++ii) {

However, the only code path that counts how many atoms in fragments[0] are local runs inside the for (ii < nfragments - 1) loop:

• deepmd-kit/source/lib/src/pairwise.cc

  Lines 99 to 109 in 73de44b

  	for (int ii = 0; ii < nfragments - 1; ++ii) {
  	// real
  	for (int jj = 0, kk = 0; jj < nqm; ++jj) {
  	if (fragments[0][jj] < nloc) {
  	forward_qmmm_map[ii * map_size + kk] = fragments[0][jj];
  	kk++;
  	}
  	if (jj == nqm - 1) {
  	nqm_real = kk;
  	}
  	}

With a single QM fragment that covers only a subset of local atoms, nloc_qm is returned as nloc instead of the local count in fragments[0]:

• deepmd-kit/source/lib/src/pairwise.cc

  Lines 147 to 150 in 73de44b

  	nloc_qm = nqm_real;
  	nloc_qmmm = nqm_real + max_fragment_real_size;
  	nall_qm = nqm;
  	nall_qmmm = nqm + max_fragment_size;

The TensorFlow op trusts that count when copying forward_qm_map, so it can read past the vector that actually contains only the QM fragment atoms:

• deepmd-kit/source/op/tf/pairwise.cc

  Lines 159 to 166 in 73de44b

  	for (int jj = 0; jj < max_nloc_qm + max_nghost_qm; ++jj) {
  	if (jj < nloc_qm[ii]) {
  	m_forward_qm_map(ii, jj) = forward_qm_maps[ii][jj];
  	} else if (jj < max_nloc_qm) {
  	m_forward_qm_map(ii, jj) = -1;
  	} else if (jj < max_nloc_qm + nghost_qm[ii]) {
  	m_forward_qm_map(ii, jj) =
  	forward_qm_maps[ii][jj - (max_nloc_qm - nloc_qm[ii])];

Impact

For frames where all non-QM entries are placeholders or where no MM fragment is present, the op can emit incorrect natoms_qm metadata and may read out of bounds while filling forward_qm_map. If single-fragment pairwise DPRc frames are not supported, the current code should reject them explicitly instead of producing inconsistent maps.

Suggested fix

Count local atoms in fragments[0] before the QMMM-fragment loop and use that value for nloc_qm. Separately, if pairwise DPRc requires at least one MM fragment, validate nfragments >= 2 and return a clear TensorFlow InvalidArgument error.

Add coverage for an idxs row with one QM fragment plus -1 placeholders, and verify nloc_qm, forward_qm_map, and padding behavior.