Weekly Progress of DP-GEN 20220912

[HuangJiameng]

Important:

Recently, some problems with model inference of Python interface in DeePMD-kit 2.1.0-2.1.3 CPU version are found. If the model is trained with the 2.1.0-2.1.3 CPU version, due to the model's flaws, the MD process may generate bad data, such as data with abnormally large forces. This issue was fixed in version 2.1.4. Please note that 2.1.0-2.1.3 GPU version is not affected. Besides, LAMMPS in 2.1.0-2.1.3 CPU/GPU version is not affected, which means if one uses LAMMPS in DeePMD-kit 2.1.0 CPU version to run MD process with a normal model, then everything is all right.

What's new

• add publication link to docs #939

Discussion

• Question about the number of 02.fp tasks (task.00x.xxxxx) #934
  Discussion on the meaning of fp_task_max, and how to control the number of candidates.
• command "model_devi_activation_func" problem #941
• Fail to run 01model_devi after 00.train #935
  model_devi jobs in param.json decide the number of iterations.

Dependency

Dpdispatcher

• add ci tests on slurm
• change source of mock-ssh-server; uncomment 3.8
• fix authing using secrets and totp

dpdata

• add option to ignore convergence check for vasp/contcar
  When extracting the results of vasp using dpdata, it automatically ignored some unconverged data, which was reasonable for the next training with deepmd. But if one only wants to extract the trajectories, some structures will be missed. Therefore, an option to ignore the convergence check and provide a warning is added.