fix dtype in PairTabOp

[njzjz]

Fix #2499.

[codecov]

Codecov Report

Patch coverage: 100.00% and no project coverage change.

Comparison is base (b13b26c) 74.40% compared to head (4712a90) 74.40%.

Additional details and impacted files

@@           Coverage Diff           @@
##            devel    #2500   +/-   ##
=======================================
  Coverage   74.40%   74.40%           
=======================================
  Files         227      227           
  Lines       23428    23428           
  Branches     1666     1680   +14     
=======================================
  Hits        17431    17431           
  Misses       4890     4890           
  Partials     1107     1107           

Impacted Files	Coverage Δ
deepmd/utils/pair_tab.py	100.00% <ø> (ø)
source/op/pair_tab.cc	94.49% <100.00%> (ø)

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

[ixbozai]

Hi, after the new release of 2.2.2, the 'use_srtab' is compatible with float precision, but new problem comes. After 20000 steps of training, all the RMSEs become NaN, and the final force field is just unusable (crush immediately when running MD). Again, this happened only for training with float precision.

[ixbozai]

step rmse_trn rmse_e_trn rmse_f_trn rmse_v_trn lr

  0      1.69e+02      1.10e+00      5.31e+00      3.45e+00    1.0e-03

10000 2.53e+01 2.92e-01 8.11e-01 7.79e-01 9.4e-04
20000 nan nan nan nan 8.9e-04
30000 nan nan nan nan 8.4e-04

[njzjz]

There are probably some problems in data stat when using pairwise tabulated potentials. The initial loss value is too large.