[Feature Request] TorchSim support for DeePMD-kit #6125

Summary

Add a long-term support for DeePMD-kit PyTorch models in TorchSim.

Detailed Description

TorchSim is a PyTorch-native atomistic simulation engine that supports batched, GPU-accelerated simulations. I developed deepmd_torchsim, a TorchSim ModelInterface wrapper for the PyTorch backend of DeePMD-kit.

• Package: https://pypi.org/project/deepmd-torchsim/
• Repository: https://github.com/rahulumrao/deepmd_torchsim
• TorchSim PR: Added DeePMD-kit PyTorch backend integration TorchSim/torch-sim#604

Current status

The wrapper supports multitask models through the head= interface. It has passed TorchSim's model tests locally using DPA-3.1-3M (22/22 tests).

Motivation

The wrapper currently is in my personal repository. In the [discussion on the TorchSim PR](TorchSim/torch-sim#604), the maintainers suggested upstreaming the integration into DeePMD-kit. They said they might consider an equivalent PR backed by long-term support from DeepModeling.

Proposal

I would like to explore the possibility of bringing deepmd_torchsim under the DeepModeling umbrella to ensure its long-term maintenance and support. Two possible approaches would be:

1. Add TorchSim support to deepmd-kit, for example, as an optional module.
2. Move deepmd_torchsim to the DeepModeling as a standalone package.

Implementation

I'm happy to help in the migration and the necessary changes to support either approach.

I would appreciate your thoughts on this and welcome any suggestions.

Further Information, Files, and Links

No response