mliap not working with recent lammps develop branch

[bernstei]

It appears that mliap is giving a very different result than symmetrix (lammps or standalone python) or pytorch MACECalculator. I'm putting together an example, but I just thought I'd warn people of this possible problem

[bernstei]

OK. This is the output of the attached scripts / lammps log files (excluding some error messages)

> egrep '^[ ]*0 [ ]*0 ' log.*
log.mliap_gpu:         0   0             -230.35339      0             -230.35339     -68778.518    
log.symmetrix_cpu:         0   0             -246.72042      0             -246.72042     -34499.435    
log.symmetrix_gpu:         0   0             -246.72042      0             -246.72042     -34499.435    

> tin 2849 : python3 compare.py 
pytorch -246.72040270810288
symmetrix -246.72042485128355

symmetrix model was created with

symmetrix_extract_mace -m ~/.cache/mace/maceomat0smallmodel -s 13 28 29 -o small.Al_Ni_Cu.symmetrix.json

mliap model was created with

mace_create_lammps_model --format mliap ~/.cache/mace/maceomat0smallmodel
mv ~/.cache/mace/maceomat0smallmodel-mliap_lammps.pt .

on a gpu node

lammps is develop head as for Nov 24, commit ac5940583
mace is develop head as of Dec 1, commit 0212de8

compare.py
config.lammps-data.txt
lammps_mliap.in.txt
lammps_symmetrix.in.txt

[bernstei]

current release branch, dated Sep 10, seems to work. I may try to bisect, but it takes forever to compile :)

[bernstei]

Looks like it broke between

lammps/lammps@3ada52d

and

lammps/lammps@02f7818

as commented in #1301 (comment)

[bernstei]

@wcwitt do we want to mark this one as a dup of #1301, or vice versa?