[BUG] Incorrect dipole prediction with LAMMPS of new version

[Yi-FanLi]

Summary

The following tasks give inconsistent results:
Task1: Compute dipole with DeePMD-kit python interface;
Task2: Compute dipole with DeePMD-kit C++ interface + LAMMPS (Oct29, 2020);
Task3: Compute dipole with DeePMD-kit C++ interface + LAMMPS (Sep20, 2021).

Task1 and Task2 give consistent predictions, but the result of Task3 is different from the others.

Deepmd-kit version, installation way, input file, running commands, error log, etc.

DeePMD-kit: v2.0.3 (commit 2223ff3), built from source.
LAMMPS versions:
Task2: commit 31f9f17
Task3: commit 584943f
The input scripts and the model are given in the attachment.
dipole_bug_reproduce.tar.gz

Task1 and Task2 predict the dipole of the first atom to be -0.0582886 0.0159309 -0.0381824, while Task3 predict the dipole of the first atom to be -0.0637354 0.0123003 -0.0416543.

Steps to Reproduce
For Task1, simply run the python script with
python run.py

The predicted dipole is in dipole.raw.

For Task2 and Task3, run the lammps task with
lmp_mpi -echo screen -in in.lammps >> output
The predicted dipole is in water.dump.

[njzjz]

I am wondering it is the same bug as #1109... Could you try to uncomment source/lmp/compute_deeptensor_atom.cpp#L73?

deepmd-kit/source/lmp/compute_deeptensor_atom.cpp

Line 73 in 159e45d

// neighbor->requests[irequest]->full = 1;

[Yi-FanLi]

I commented Line 73 and recompiled DeePMD-kit and LAMMPS. Now it gives the consistent results with the other test cases.