use lammps inference not quickly, but use python to run model is fast

[WorldPeace123]

in pair_allegro_kokkos.cpp line 296. the model need input ,after compute give the lmp an output.
I test the " auto output = this->model.forward(input_vector).toGenericDict();" time ,
in RTX3060, 1step need 0.4s.and i write a python code .load the model and give input.,and 1 step need 0.05s!!!
so,what is the problem in lammps?and how to make the model inference fast?
Thanks !

[WorldPeace123]

1: as we all know ,use scriptmodel run in C++ have good performance.but in this model ,it seemd bad.
2:netron can see the model input shape,such as resnet50,but deployed.pth cant see in netron.if i get the true input shape ,i can use lot of tools to make the model run fast ,but scriptmodel is different. so if we change the model save ,maby good?
3:now have many ai compile tools,this model cant let ai compiler to change that have better performance :-(

[Linux-cpp-lisp]

The TorchScript compiler can take a few iterations to warm up, which often manifests in very slow time steps initially.

[WorldPeace123]

TorchScript 编译器可能需要几次迭代来预热，这通常在最初表现为非常慢的时间步长。
i know my friend,but in pair_allegro.cpp ,auto output = this->model.forward（input_vector）.toGenericDict（）The average time taken is 0.4s after the warm-up mechanism.

i get the log ,can you help me to know what reason make the C++ run model dont not fast than use python to run model?
Thanks !!!

this is python
##jit.load.mod and use
/home/worldpeace/anaconda3/envs/tvm/lib/python3.11/site-packages/nequip/init.py:20: UserWarning: !! PyTorch version 2.5.1 found. Upstream issues in PyTorch versions 1.13.* and 2.* have been seen to cause unusual performance degredations on some CUDA systems that become worse over time; see mir-group/nequip#311. The best tested PyTorch version to use with CUDA devices is 1.11; while using other versions if you observe this problem, an unexpected lack of this problem, or other strange behavior, please post in the linked GitHub issue. warnings.warn(
Time :0.6278s
Time :0.5497s
Time :1.6636s
Time :0.3279s
Time :0.0221s
Time :0.0236s
Time :0.0239s
Time :0.0243s
Time :0.0231s
Time :0.0288s

this is pair_allegro
Per MPI rank memory allocation (min/avg/max) = 5.31 | 5.31 | 5.31 Mbytes Step Time PotEng KinEng TotEng Temp Press Volume Density
0 0 -5115.3514 247.66244 -4867.6889 1000 124912.37 35001.599 9.8101743
model.forward Time is : 0.445529 stoTensor().cpu Time is : 0.000252 s
Pair All Time is : 0.452624 s
model.forward Time is : 3.49675 s
toTensor().cpu Time is : 7.3e-05 s
Pair All Time is : 3.50579 s
model.forward Time is : 0.283893 s
toTensor().cpu Time is : 0.000103 s
Pair All Time is : 0.286683 s
model.forward Time is : 0.282643 s
toTensor().cpu Time is : 6e-05 s
Pair All Time is : 0.290819 s
model.forward Time is : 0.283949 s
toTensor().cpu Time is : 0.000106 s
Pair All Time is : 0.28688 s
model.forward Time is : 0.282495 s
toTensor().cpu Time is : 5.4e-05 s
Pair All Time is : 0.286974 s
model.forward Time is : 0.282829 s
toTensor().cpu Time is : 0.000113 s