Release 1.0.0
Summary
This is the first release of NVIDIA NIM for Batched Geometry Relaxation.
Features
The 1.0.0 release includes the following key features:
- Multiple MLIP Models: Support for MACE, AIMNet2 (with NSE variant), and TensorNet models.
- Dynamic Batching: Automatic batch size estimation optimized for available GPU memory.
- GPU-accelerated FIRE2 Optimizer: High-performance geometry optimization directly on GPU.
- Cell Optimization: Optional unit cell optimization with pressure tolerance constraints.
- DFT-D3 Dispersion: Optional DFT-D3(BJ) dispersion corrections for improved van der Waals interactions.
- Per-request Parameters: Override global optimization parameters on a per-request basis.
- Active Masks: Freeze specific atoms during optimization.
Limitations
The following limitations apply to the 1.0.0 release:
- External model weights: You must provide AIMNet2 and TensorNet models externally by using a volume mount. Only MACE is downloaded automatically. For instructions, refer to Custom Models.
- Fixed PBC mode: Periodic boundary conditions (PBC) mode is fixed at container startup. You cannot mix periodic and non-periodic structures in the same container instance.
- Memory constraints: Very large structures can exceed GPU memory limits and result in an out-of-memory HTTP error. Use the
/v1/statusendpoint to check themax_system_size(maximum number of atoms per system) for your configuration.
Example: Checking the Maximum System Size
To view the constraints for your deployment, query the /v1/status endpoint:
For more details, refer to the API Reference.