ml-mlip-speed

What problem does it solve?

Selecting an MLIP for atomistic simulations is hard because inference latency and memory footprint vary widely by model and system size, which can make planned MD runs too slow or crash with OOM errors.

Core Features & Use Cases

• Inference speed benchmarking: Measures per-atom inference time by running short NVE MD steps on NaCl supercells across increasing sizes.
• Peak memory/VRAM estimation: Tracks peak memory usage during inference (GPU VRAM via PyTorch or host RAM via psutil).
• Multi-environment model coverage: Supports benchmarking models that require separate Conda environments by incrementally updating a shared speed_benchmark.yaml.
• Plot consolidation: Regenerates comparative speed and memory plots from accumulated YAML results using a single flag for final visualization.
• Use Case: Compare MACE, MatGL, and FAIRCHEM checkpoints to pick a model tier (fast/low-cost vs heavier/accurate) for long MD on large cells.

Quick Start

Run the benchmark script with models and providers specified so it produces speed_benchmark.yaml and per-hardware plots in your chosen output directory.