Exception: expected scalar type Double but found Float

[duneandre]

Hi,

I have trained a Machine Learning Potential using NequIP version 0.5.6.

Now I want to run MD with this MLP using LAMMPS (version 29 August 2024) and pair_nequip (latest git version). I am running it on another machine on which I have installed NequIP version 0.6.1. (the latest). The calculation did not work and I received the following error :
Exception: expected scalar type Double but found Float

Do you know how to solve this problem?

Thanks for your help !

[fxcoudert]

I am encountering this issue as well:

• a model trained and deployed with NequIP version 0.5.6
• trying to run on LAMMPS with latest pair_nequip from git main branch

I have found and read the previous issue at #51 which seems similar. It states that the problem should be fixed after pull request #52 from @anjohan but I am still seeing the issue with latest git version (which includes this commit).

Is there any configuration that can be tweaked to pair_nequip? Or any advice on how to redeploy the model?

[baham2]

Hello, I was wondering if you were able to solve this issue? I am also facing the "Exception: expected scalar type Double but found Float" error, with the following versions: (default_dtype: float64, model_dtype: float32 in config.yaml)
nequip 0.6.0
pair_nequip 0.5.1
lammps stable_29Sep2021_update2

I tried to solve it by converting the pth file into float32 based on #21 (However, using torch.jit as I got a serialization error without it), but I got the following error:
The indicated TorchScript file does not appear to be a deployed NequIP model; did you forget to run nequip-deploy? (src/pair_nequip.cpp:181)

[afour9961207]

Hi all,

I'm experiencing a similar issue. The Lammps is compiled successfully in my personal PC and cluster, but when I try to use Nequip in LAMMPS, it fails.

Environment Details:
nequip 0.5.6
pair_nequip 0.5.2
Torch 1.12.1+cu11.6
lammps-stable_23Jun2022_update4

I've tried generating different potential.pth files by modifying the YAML configuration and retraining:

1. default_dtype: float32
2. default_dtype: float64
3. default setting like neq 0.6.0 (default_dtype: float64 model_dtype: float32 allow_tf32: true)

Each test results in a core dump, with the following errors:

1. terminate called after throwing an instance of 'c10::Error'
   what(): expected scalar type Double but found Float
2. RuntimeError: expected scalar type Float but found Double
3. terminate called after throwing an instance of 'c10::Error'
   what(): expected scalar type Double but found Float

My Lammps input is

units metal
dimension 3
newton off
boundary p p p
box tilt large
atom_style charge
neighbor 2.0 bin
read_data structure_atom.dat.q
pair_style nequip
pair_coeff * * ./potential.pth Sr Ti O
run 0

And the structure_atom.dat.q is just supercell of SrTiO3 containing 60 atoms

I'm quite confused about what the issue might be. Any help would be greatly appreciated! Thanks in advance.

[Linux-cpp-lisp]

Hi all---this is indeed strange that you are still seeing this issue with what should have been a fix. I'm not sure why... we should have some more patches out within some weeks that should resolve this; thanks for your patience.