[Feat]: MD Accelaration and LAMMPS Integration

[HNUSTLingjunWu]

Contact Details

lingjun.wu@hotmail.com

Feature Description

Dear Developers:

I've tried run MD using Mattersim via ASE, however I found the speed is quite slow, I'm just wondering if there is anyway to accelarate the simulation? By the way, can I integrate Mattersim with LAMMPS? thanks so much.

Best wishes,
Lingjun

Motivation

1. To accelarate MD simulation;
2. To integrate with LAMMPS, which is quite fast.

Proposed Solution

No response

Contribution Interest

• I'm interested in potentially implementing this feature
• I can provide guidance, but cannot implement
• I'm just suggesting the feature for the community

Code of Conduct

• I agree to follow the project's Code of Conduct

[luzihen]

Lingjun,

What hardware are you running the model on? The speed should be comparable to most GNN-based MLFFs.

We do have a LAMMPS interface. Just did not get the bandwidth to clean up the code and push that to the repo. So, while we welcome contribution from the community, for this case, I'd suggest wait a bit so as not to duplicate the work.

PS, interfacing with LAMMPS likely won't accelerate the simulation as the model inference bit is the most time-consuming part.

[HNUSTLingjunWu]

Dear Ziheng, The mattersim is a wonderful model however I've encountered some problems and would be really grateful if you can give me some advice. I'm using RTX3090, my system contains 3000 atoms and I've tested that it tooks about 4 hours to run 50 ps NVT using ASE, which would be too slow as I wish to inspect MSD, which may need simulation of about 10 ns, no idea if there is anyway I can accelerate the simulation? thanks so much. Best wishes, Lingjun

…

________________________________ From: Ziheng LU ***@***.***> Sent: 20 December 2024 14:08 To: microsoft/mattersim ***@***.***> Cc: HNUSTLingjunWu ***@***.***>; Author ***@***.***> Subject: Re: [microsoft/mattersim] [Feat]: MD Accelaration and LAMMPS Integration (Issue #66) Lingjun, What hardware are you running the model on? The speed should be comparable to most GNN-based MLFFs. We do have a LAMMPS interface. Just did not get the bandwidth to clean up the code and push that to the repo. So, while we welcome contribution from the community, for this case, I'd suggest wait a bit so as not to duplicate the work. PS, interfacing with LAMMPS likely won't accelerate the simulation as the model inference bit is the most time-consuming part. — Reply to this email directly, view it on GitHub<#66 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BG2DYDGBTYBMMJ5E7IIHW4L2GQQFTAVCNFSM6AAAAABT6UUW3WVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNJXGA4DMNZSGM>. You are receiving this because you authored the thread.

[JonathanSchmidt1]

Does the lammps interface include possibilities for running larger systems on multiple gpus?
Right now even with H100 the system size is quite limited.
Or do you have any other way to do e.g. offloading from the gpu to cpu to increase systems size.

edit: wrapping the energy calculation in the forward call with torch.autograd.graph.save_on_cpu(pin_memory=True) allows to go to significantly larger systems while still being faster than cpu.

[HNUSTLingjunWu]

Dear Jonathan, Do you mean to use CPU instead of GPU can accelerate the simulation? Best wishes, Lingjun

…

________________________________ From: Jonathan Schmidt ***@***.***> Sent: 06 January 2025 11:39 To: microsoft/mattersim ***@***.***> Cc: HNUSTLingjunWu ***@***.***>; Author ***@***.***> Subject: Re: [microsoft/mattersim] [Feat]: MD Accelaration and LAMMPS Integration (Issue #66) Does the lammps interface include possibilities for running larger systems on multiple gpus? Right now even with H100 the system size is quite limited. Or do you have any other way to do e.g. offloading from the gpu to cpu to increase systems size. — Reply to this email directly, view it on GitHub<#66 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/BG2DYDBEL4SKJKVCBBYC7N32JJTPXAVCNFSM6AAAAABT6UUW3WVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNZSHEZTKMJXHE>. You are receiving this because you authored the thread.Message ID: ***@***.***>

[JonathanSchmidt1]

Dear Lingjun,

Thank you for your response.

I have two questions:

1. Does your LAMMPS interface support running system that do not fit into the vram of one gpu on multiple GPUs using MPI?
2. Offloading activations from VRAM to RAM or using pipelining allows for training larger models. In the case of force fields, such techniques could also enable simulations of larger systems, as force calculations for the M3GNet model require backpropagation. If your code does not currently support multiple GPUs, have you explored any offloading strategies to scale to larger systems?

Specifically, this approach keeps calculations on the GPU while transferring intermediate activations to RAM or recomputing them during backpropagation for force calculations. After my initial inquiry, I experimented with a straightforward strategy in PyTorch—wrapping the energy computation during the forward pass with torch.autograd.graph.save_on_cpu(pin_memory=True). This indeed allowed me to run significantly larger systems at an acceptable speed, which was still faster than CPU-based computations.

best wishes,
Jonathan