ml-mlip-speed

You can benchmark MLIP inference speed and VRAM by executing short NVE molecular dynamics runs on NaCl supercells. This process measures per-atom inference latency and peak memory consumption, persisting the results into a YAML file for comparison.

To compare MACE, MatGL, and FAIRCHEM memory consumption, you run short MD steps across increasing NaCl supercell sizes. The benchmark tracks peak GPU VRAM via PyTorch or host RAM via psutil to guide model selection.

Yes, you can benchmark machine learning interatomic potentials across separate Conda environments. The script incrementally updates a shared speed_benchmark.yaml file, allowing you to aggregate results from different model providers without conflicts.

You can regenerate comparative speed and memory plots without rerunning MD simulations by using a specific plot consolidation flag. This reads the accumulated results from your existing speed_benchmark.yaml file to produce final visualizations.

Your chosen MLIP crashes with OOM errors because inference latency and memory footprint vary widely by model and system size. Benchmarking peak memory consumption across different supercell sizes helps prevent these out-of-memory crashes during long MD runs.

The best way to select an MLIP model tier for long MD on large cells is to benchmark per-atom latency and peak VRAM. Comparing results across providers like MACE, MatGL, and FAIRCHEM helps balance fast low-cost models against heavier accurate ones.