[BUG] Run MD with LAMMPS has a temperature instability problem on the AMD ROCM platform

[LLfbforever]

Bug summary

When executing the calculation example, the temperature is unstable when LAMMPS runs the force field trained by deepmd

DeePMD-kit Version

master branch

TensorFlow Version

tensorflow2.7.0

How did you download the software?

Offline packages

Input Files, Running Commands, Error Log, etc.

in file to add the command line：
plugin load libdeepmd_lmp.so

The output result of lammps operation:
the temperature value is unstable. When temp=2000k, more than 10,000 temp values will appear during the operation.

Steps to Reproduce

1. install deepmd-kit
2. install lammps
3. run
4. Draw a graph of temperature change

Further Information, Files, and Links

No response

[LLfbforever]

by finding bugs:
Through debugging, locate the calculation process of the TabulateFusionSeAGrad operator in deepmd-kit. After modifying the number of iterations of the second for loop of kernel: tabulate_fusion_se_a_grad_fifth_order_polynomial in the /deepmd-kit-master/source/lib/src/rocm/tabulate.hip.cu file.
Namely:
(1) The for loop should start from warp_idx→nnei
(2) Change int ii = 0 in the for loop to int ii = warp_idx, and all ii+warp_idx in the for loop are changed to ii
Running the study again, the temperature reached a steady state.

[njzjz]

Fixed in #2357.