ML-IAP model not loading to LAMMPS #556
Answered
by
cw-tan
apoletayev
asked this question in
Q&A
-
|
A next-step error to a preceding question (#555). With updates to torch=2.8.0 and nequip=0.15.0 and OEQ on an H100 the models train but do not load into LAMMPS. Same versions of everything for training, preparing, and LAMMPS. I would appreciate any help or pointers on what could be the origins of these errors. Error: I then suppressed this with the environment variable If a model is prepared for ML-IAP with the OEQ modifier, the error is again different: The two errors are the same irrespective of whether the model was trained with OEQ; they only differ by whether the OEQ was included in the ML-IAP preparation step. In case it makes any difference, LAMMPS was built with the GPU package rather than KOKKOS because for a very heterogeneous cluster KOKKOS would have to be recompiled for every GPU generation (is KOKKOS necessary for ML-IAP?). I am not sure why the error trace mentions the CPU backend because everything was compiled and built and run from a GPU. LAMMPS cmake call: |
Beta Was this translation helpful? Give feedback.
Answered by
cw-tan
Oct 28, 2025
Replies: 1 comment 7 replies
-
|
Hi @apoletayev , Our LAMMPS ML-IAP integration was implemented with Kokkos in mind. I'm guessing that some of the errors are related. I'm not sure about the serialization error, but indicates that the I think you should compile LAMMPS ML-IAP with Kokkos. I'm not sure entirely about this, but my impression is that the GPU generation constraint for Kokkos you have in mind might not actually be that strict. @wcwitt and @anjohan would know better. Side note that training with OEQ or not should be independent of whatever happens at inference since the OEQ model modification is not persistent, i.e. we modify the model for training, but when it's reconstructed during |
Beta Was this translation helpful? Give feedback.
7 replies
-
I believe KOKKOS is required. If I remember correctly, the LAMMPS updates providing the possibility for |
Beta Was this translation helpful? Give feedback.
-
|
Unfortunately, I am getting similar errors with a kokkos-containing build of LAMMPS. call to cmake: call to lammps (29Aug2024_update4): model loading in LAMMPS: Error trace (model w/o OEQ): Error trace (model with OEQ): The errors look similar to before switching to kokkos ... does this also need some minimum release of LAMMPS (I'm using 29Aug2024_update4) or a minimum version of KOKKOS (mine is 4.3.1)? |
Beta Was this translation helpful? Give feedback.
-
|
I believe one needs the update specified in this blogpost: https://developer.nvidia.com/blog/enabling-scalable-ai-driven-molecular-dynamics-simulations/ , specifically
|
Beta Was this translation helpful? Give feedback.
Answer selected by
apoletayev
-
|
This works well! Finally a 0.15.0 simulation with OEQ for me! The One last question - I was using I thought building neighbor lists on the GPU was part of the point of this integration, is this something to worry about? This seems to only pertain to minimization, not dynamics. |
Beta Was this translation helpful? Give feedback.
-
|
Hm, we've seen this a few times recently. I suspect it's something specific to minimization with Kokkos, unrelated to NequIP, but not 100% sure. Please let us know if any results look suspicious, otherwise I think we have to live with any performance hit for now (hopefully small). I'll double check with LAMMPS developers as well. |
Beta Was this translation helpful? Give feedback.
-
|
LAMMPS developers confirmed this shouldn't affect performance significantly, perhaps not at all. Those warnings are completely unrelated to the pair style, and they will likely go away in future versions of LAMMPS. |
Beta Was this translation helpful? Give feedback.
-
|
Thank you both! Looking forward to doing more science with nequip. |
Beta Was this translation helpful? Give feedback.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I believe one needs the update specified in this blogpost: https://developer.nvidia.com/blog/enabling-scalable-ai-driven-molecular-dynamics-simulations/ , specifically