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

Problem

The NVNMD env-mat TensorFlow ops accept tensor-provided neighbor lists (nei_mode == 4) but then ignore the neighbor-list data and rebuild neighbors from coordinates.

Both NVNMD env-mat paths classify mesh.size() > 16 as a tensor-stored neighbor list:

• deepmd-kit/source/op/tf/prod_env_mat_multi_device_nvnmd.cc

  Lines 410 to 427 in 73de44b

  	int nei_mode = 0;
  	bool b_nlist_map = false;
  	if (mesh_tensor.shape().dim_size(0) == 16) {
  	// lammps neighbor list
  	nei_mode = 3;
  	} else if (mesh_tensor.shape().dim_size(0) == 6) {
  	// manual copied pbc
  	assert(nloc == nall);
  	nei_mode = 1;
  	b_nlist_map = true;
  	} else if (mesh_tensor.shape().dim_size(0) == 0) {
  	// no pbc
  	assert(nloc == nall);
  	nei_mode = -1;
  	} else if (mesh_tensor.shape().dim_size(0) > 16) {
  	// pass neighbor list inside the tensor
  	nei_mode = 4;
  	} else if (mesh_tensor.shape().dim_size(0) == 7 ||

• deepmd-kit/source/op/tf/prod_env_mat_multi_device_nvnmd.cc

  Lines 662 to 679 in 73de44b

  	int nei_mode = 0;
  	bool b_nlist_map = false;
  	if (mesh_tensor.shape().dim_size(0) == 16) {
  	// lammps neighbor list
  	nei_mode = 3;
  	} else if (mesh_tensor.shape().dim_size(0) == 6 ||
  	mesh_tensor.shape().dim_size(0) == 7) {
  	// manual copied pbc
  	nei_mode = 1;
  	b_nlist_map = true;
  	} else if (mesh_tensor.shape().dim_size(0) == 0 ||
  	mesh_tensor.shape().dim_size(0) == 1) {
  	// no pbc
  	nei_mode = -1;
  	} else if (mesh_tensor.shape().dim_size(0) > 16) {
  	// pass neighbor list inside the tensor
  	nei_mode = 4;
  	} else {

However, the shared NVNMD _prepare_coord_nlist_cpu() treats every mode except 3 as "build nlist by myself":

• deepmd-kit/source/op/tf/prod_env_mat_multi_device_nvnmd.cc

  Lines 231 to 276 in 73de44b

  	template <typename FPTYPE>
  	static void _prepare_coord_nlist_cpu(OpKernelContext* context,
  	FPTYPE const** coord,
  	std::vector<FPTYPE>& coord_cpy,
  	int const** type,
  	std::vector<int>& type_cpy,
  	std::vector<int>& idx_mapping,
  	deepmd::InputNlist& inlist,
  	std::vector<int>& ilist,
  	std::vector<int>& numneigh,
  	std::vector<int*>& firstneigh,
  	std::vector<std::vector<int>>& jlist,
  	int& new_nall,
  	int& mem_cpy,
  	int& mem_nnei,
  	int& max_nbor_size,
  	const FPTYPE* box,
  	const int* mesh_tensor_data,
  	const int& nloc,
  	const int& nei_mode,
  	const float& rcut_r,
  	const int& max_cpy_trial,
  	const int& max_nnei_trial) {
  	inlist.inum = nloc;
  	if (nei_mode != 3) {
  	// build nlist by myself
  	// normalize and copy coord
  	if (nei_mode == 1) {
  	int copy_ok = _norm_copy_coord_cpu(coord_cpy, type_cpy, idx_mapping,
  	new_nall, mem_cpy, *coord, box, *type,
  	nloc, max_cpy_trial, rcut_r);
  	OP_REQUIRES(context, copy_ok,
  	errors::Aborted("cannot allocate mem for copied coords"));
  	*coord = &coord_cpy[0];
  	*type = &type_cpy[0];
  	}
  	// build nlist
  	int build_ok = _build_nlist_cpu(ilist, numneigh, firstneigh, jlist,
  	max_nbor_size, mem_nnei, *coord, nloc,
  	new_nall, max_nnei_trial, rcut_r);
  	OP_REQUIRES(context, build_ok,
  	errors::Aborted("cannot allocate mem for nlist"));
  	inlist.ilist = &ilist[0];
  	inlist.numneigh = &numneigh[0];
  	inlist.firstneigh = &firstneigh[0];
  	} else {

The non-NVNMD implementation has a separate nei_mode == 4 branch that copies ilist, numneigh, and jlist from the mesh tensor:

• deepmd-kit/source/op/tf/prod_env_mat_multi_device.cc

  Lines 1471 to 1506 in 73de44b

  	inlist.inum = nloc;
  	if (nei_mode != 3 && nei_mode != 4) {
  	// build nlist by myself
  	// normalize and copy coord
  	if (nei_mode == 1) {
  	int copy_ok = _norm_copy_coord_cpu(coord_cpy, type_cpy, idx_mapping,
  	new_nall, mem_cpy, *coord, box, *type,
  	nloc, max_cpy_trial, rcut_r);
  	OP_REQUIRES(context, copy_ok,
  	errors::Aborted("cannot allocate mem for copied coords"));
  	*coord = &coord_cpy[0];
  	*type = &type_cpy[0];
  	}
  	// build nlist
  	int build_ok = _build_nlist_cpu(ilist, numneigh, firstneigh, jlist,
  	max_nbor_size, mem_nnei, *coord, nloc,
  	new_nall, max_nnei_trial, rcut_r);
  	OP_REQUIRES(context, build_ok,
  	errors::Aborted("cannot allocate mem for nlist"));
  	inlist.ilist = &ilist[0];
  	inlist.numneigh = &numneigh[0];
  	inlist.firstneigh = &firstneigh[0];
  	} else if (nei_mode == 4) {
  	std::memcpy(&ilist[0], 16 + mesh_tensor_data, sizeof(int) * nloc);
  	std::memcpy(&numneigh[0], 16 + nloc + mesh_tensor_data, sizeof(int) * nloc);
  	for (int ii = 0, kk = 0; ii < nloc; ++ii) {
  	jlist[ii].resize(numneigh[ii]);
  	std::memcpy(&jlist[ii][0], 16 + 2 * nloc + kk + mesh_tensor_data,
  	sizeof(int) * numneigh[ii]);
  	firstneigh[ii] = &jlist[ii][0];
  	kk += numneigh[ii];
  	}
  	inlist.ilist = &ilist[0];
  	inlist.numneigh = &numneigh[0];
  	inlist.firstneigh = &firstneigh[0];
  	} else {

Impact

Calls to ProdEnvMatANvnmdQuantize or ProdEnvMatAMixNvnmdQuantize with a tensor-stored neighbor list can silently compute descriptors using a rebuilt distance neighbor list instead of the caller-provided list. That is especially risky for tests or integrations that intentionally provide a prefiltered or otherwise non-distance-equivalent neighbor list.

Suggested fix

Mirror the non-NVNMD nei_mode == 4 branch in the NVNMD helper, including the ilist, numneigh, jlist, and firstneigh setup.

Add CPU coverage for both NVNMD env-mat ops where the tensor-provided neighbor list differs from the distance-built neighbor list, and verify the op follows the tensor list.