Reproducibility and determinism

MLIP inference on a GPU is not guaranteed to be bit-reproducible by default: parallel reductions can accumulate in a hardware- and software-dependent order. Assess numerical sensitivity on the target backend, model, hardware, and software stack.

When exact comparison is required, use --deterministic to request deterministic algorithms, then verify the produced artifacts on the complete target stack.

--deterministic

--deterministic is accepted by every compute subcommand (opt, tsopt, freq, irc, scan, scan2d, scan3d, path-opt, path-search, all, sp). It turns on torch.use_deterministic_algorithms and an index_reduce_ shim for operations controlled by mlmm-toolkit. PyTorch raises if a selected operation lacks a deterministic implementation.

mlmm opt -i complex.pdb --parm enzyme.parm7 -q 0 --deterministic
mlmm all -i r_complex.pdb p_complex.pdb -c PRE -q -1 --deterministic

• It is process-global: setting it on all propagates to every internal stage; you do not pass it per stage.

• It can change performance because deterministic and default kernels may use different implementations.

• For PyTorch operations under its control, an unavailable deterministic implementation raises instead of silently using a known nondeterministic implementation. Backend SDK and custom operations still require separate verification.

• The environment variable MLMM_STRICT_DETERMINISTIC=1 is the equivalent entry point for CI or the direct Python API.

Backend support

ML backend	--deterministic
uma	deterministic mode accepted; verify the installed model/SDK on the target system
orb	deterministic mode accepted; verify the installed model/SDK on the target system
mace	deterministic mode accepted; verify the installed model/SDK on the target system
aimnet2	not supported — rejected (see below)
custom (--calc-file)	not supported — rejected because the supplied calculator is outside mlmm-toolkit’s control

The default hessian_ff MM low-level layer runs on CPU; OpenMM can instead use its selected device, including CUDA. Exact end-to-end comparison still requires fixed inputs (including the topology), software versions, hardware, and backend configuration.

Precision and reproducibility

Running in --precision fp64 changes numerical precision but does not by itself guarantee bit-identical GPU execution. --deterministic requests deterministic algorithms; confirm exact reproducibility for the complete target stack.

--precision fp64 and Hessian storage precision (H_double, default true) are independent knobs; supported fp32 configurations can set H_double: false, while passing --precision fp64 forces the Hessian to fp64 so the optimizer linear algebra cannot silently run in a lower precision than the model.

Precision defaults are backend-specific (UMA fp32; ORB and MACE fp64). For selection guidance and performance tradeoffs, see Device & HPC Setup → Backend precision defaults.

AIMNet2 limitations

AIMNet2 does not support these features:

• --precision fp64 — AIMNet2’s model inputs are cast to float32 upstream, so an “fp64” run would not actually be fp64.

• --deterministic — AIMNet2 computes forces through a custom CUDA kernel that lies outside torch.use_deterministic_algorithms control, so the flag cannot enforce exact force repeatability. PyTorch’s deterministic mode neither detects nor controls the custom op, so the limitation is reported explicitly.

UMA, Orb, and MACE accept the flag; verify exact repeatability for the installed backend/model/SDK and target stack.