FastMACE

The fastest way we know to run one long MACE trajectory on one NVIDIA GPU, exact to MACE within float32 rounding, checked on your structure.

FastMACE runs molecular dynamics with pretrained MACE models (MH-1, MP-0, OMAT-0, MatPES, MACE-POLAR-1, OMOL-0, MACE-OFF). It owns the whole MD loop: CUDA graphs over each step, exact model rewrites, a cuEquivariance and OpenEquivariance hybrid backend, and a planner that picks the fastest engine that passes its checks on your structure.

import fastmace as fm

run = fm.md("NaCl.cif", model="mh1", head="matpes_r2scan", ensemble="NVT",
            temperature_K=300, steps=1_000_000, repeat=(3, 3, 3), output="run/", seed=1)
print(run.summary)

Documentation: user guide and API reference. Start with docs/quickstart.md, or build the site locally with mkdocs build --strict.

Status (25 October 2026)

Version 1.0.0. Phases 0 to 5 are accepted: extraction of the research engine, contracts, safety (gates, fallback, durable restarts, provenance), MC and the remaining integrators, D3, relaxations and the planner with runtime adaptation. Phase 6's packaging, documentation, examples and clean-install acceptance checks passed, including the prebuilt hybrid backend on Python 3.13. See the acceptance report for qualification limits and known findings. The design is in SPEC.md, progress in plans/README.md.

Why FastMACE

Measured by us on an RTX 5060 Ti, float32, 216 to 1,944 atom cells (TorchSim comparison, positioning):

• 3.8 to 10 times faster per step than TorchSim 0.6.2, and 1.1 to 2 times faster than Symmetrix-XL in LAMMPS; 3 to 22 times faster than MACE's ASE calculator. MH-1 on 216-atom NaCl: 2.90 ms per 1 fs step. The margin is largest on small cells; we claim nothing beyond the measured sizes and GPUs.
• Exact and checked. Every rewrite is exact to rounding and is checked against the original MACE module at setup. An optional gate checks the fast path on your own structure and falls back safely if it fails.
• Species swaps at graph speed. MC swaps and transmutations are transactional events inside the captured graph (experimental).
• MH-1 and PolarMACE at graph speed, with upstream bug fixes.
• D3(BJ) on the GPU, in MD, relaxations and quenches.
• Complete restarts: RNG state, exchanged species, schedule position, model hash, plan and versions.
• pip install, no nvcc, no LAMMPS build.

Use something else when

• You have many small independent structures (screening, batched relaxations): use a batched engine such as TorchSim or NVIDIA ALCHEMI. FastMACE runs one structure at a time.
• Your cell has more than about 100,000 atoms, or you need several GPUs: use MACE's LAMMPS ML-IAP interface or Symmetrix-XL, which decompose the domain and use less memory per atom.
• Your model is not MACE 0.3.x (MatterSim, UMA, SevenNet, ORB, NequIP ...): TorchSim or ALCHEMI run many model families.
• You run on AMD GPUs or CPUs: Symmetrix runs there; FastMACE is NVIDIA only.
• You need LAMMPS's ecosystem (PLUMED, enhanced sampling, its many fixes): use LAMMPS.

Install

pip install torch==2.14.1 --index-url https://download.pytorch.org/whl/cu130
pip install cuequivariance-ops-torch-cu13
pip install "fastmace[full]"
fastmace doctor

FastMACE 1.0 supports torch 2.14.1+cu130 and Python 3.10 to 3.13 on Linux. Extras: [oeq] (OpenEquivariance hybrid backend), [polar] (PolarMACE), [d3] (torch-dftd reference), [full] (all three), [dev] and [docs]. For development: pip install -e ".[dev]".

Licence

FastMACE is MIT, see LICENSE and NOTICE.md. Model weights keep their own licences; many MACE foundation models, MH-1 among them, are under the Academic Software License (non-commercial). See docs/licences.md.