Metatomic backends (fat, RGPOT, ASE)

Three ways to evaluate the same Metatomic model (for example PET-MAD) from eOn / pyeonclient:

Path	How	Torch linked into host?	Use when
Fat / native	potential = Metatomic / make_backend("metatomic")	Yes (-Dwith_metatomic)	Default, conda-forge, lowest overhead
RGPOT / engine	potential = RGPOT, backend = metatomic / make_backend("rgpot_metatomic")	No on thin hosts; torch lives in libmetatomic_engine.so	Optional plugin on base wheels
ASE wrap	metatomic_ase.MetatomicCalculator → make_backend("ase") or make_backend("ase_metatomic")	Python metatomic-torch only	Cookbook / ASE workflows; same energies

conda-forge distributes the fat native path. The RGPOT engine is the thin-host path. ASE is the Python calculator path.

Benchmark (PET-MAD)

Single-point energy and force on a 14-atom PET-MAD geometry (pet-mad-s-v1.5.0.pt, CPU). Fat and RGPOT-dlopen match to machine precision; ASE agrees within ~10 µeV. Fat and RGPOT are ~10× faster per call than the ASE wrap on this workload (the ASE path pays Python / neighbor-list setup each evaluation).

Numbers and figure are generated at docs build time from the committed JSON (../fig/data/metatomic_backend_bench.json). Commit the JSON; leave the SVG/PNG to sphinx-build via scripts/plot_metatomic_backend_bench.py.

pyeonclient metatomic backends on PET-MAD s v1.5 (CPU, 14 atoms). Left: wall time per force call. Right: energy relative to fat (µeV).

Backend	Energy (eV)	max |F| (eV/Å)	ms / call
fat (metatomic)	-71.980186	37.64625	21.7
RGPOT dlopen	-71.980186	37.64625	22.6
ASE wrap	-71.980194	37.64625	206.1

Timings depend on host CPU, torch build, and warm vs cold neighbor lists. The plot shows packaging cost on this workload to order of magnitude.

Note

Fat C++ Metatomic and the Python metatomic-torch extension both register TORCH_LIBRARY(metatomic). Loading both in one process aborts. Compare them in separate processes (the compare script does this), or use a base pyeonclient build (no -Dwith_metatomic) when wrapping ASE calculators in the same interpreter as make_potential_from_ase.

pyeonclient key → factory

import pyeonclient as pyec
from pyeonclient.backends import make_backend, list_backends

print(list_backends())
# ['ase', 'ase_metatomic', 'lj', 'metatomic', 'metatomic_dlopen', ...]

# Fat (needs -Dwith_metatomic=true)
pot = make_backend("metatomic", model_path="pet-mad.pt", device="cpu")

# RGPOT + dlopen engine
pot = make_backend(
    "rgpot_metatomic",
    model_path="pet-mad.pt",
    engine_path="libmetatomic_engine.so",
    device="cpu",
)

# ASE MetatomicCalculator (PET-MAD-safe load)
pot = make_backend("ase_metatomic", model_path="pet-mad.pt", device="cpu")
# equivalent:
# from pyeonclient.backends import make_metatomic_ase_calculator
# pot = make_backend("ase", calculator=make_metatomic_ase_calculator("pet-mad.pt"))

m = pyec.Matter(pot, pyec.Parameters())

make_backend("ase_metatomic", …) applies a small load compatibility shim for exported models (including PET-MAD) where upstream load_atomistic_model re-wraps a scripted AtomisticModel and misses _model_capabilities_outputs_names. Prefer that helper over raw MetatomicCalculator(load_atomistic_model(path)) on PET-MAD.

Reproducible bench (pixi)

From a clean tree (builds fat + ASE-safe pyeonclient, runs the three-way compare, writes the JSON and regenerates the figure):

# PET-MAD + d016_pos.con live under subprojects/gpr_optim/bench_data/petmad/
# (rsynced from ../gpr_optim when missing). Override with EON_PET_MAD_*.
pixi run -e mta-bench mta-backend-bench

Task	What it does
mta-backend-bench	meson fat + ASE cores, compare, write JSON + SVG
mta-backend-bench-skip-build	re-run compare/plot only (EON_MTA_BENCH_SKIP_BUILD=1)
mta-backend-plot	plot from committed JSON only (no C++ build)

Build trees: bbdir-mta-bench/ (fat Metatomic + RGPOT engine) and bbdir-mta-bench-ase/ (pyeonclient built without C++ metatomic, so ASE metatomic-torch alone owns TORCH_LIBRARY(metatomic)).

# optional overrides
export EON_PET_MAD_MODEL=/path/to/pet-mad-s-v1.5.0.pt
export EON_PET_MAD_POS=/path/to/pos.con
export RGPOT_METATOMIC_ENGINE=/path/to/libmetatomic_engine.so
pixi run -e mta-bench mta-backend-bench-skip-build

# figure alone (docs build also regenerates from JSON)
pixi run -e mta-bench mta-backend-plot
# or: sphinx-build docs/source docs/build/html

Driver: scripts/run_metatomic_backend_bench.sh. Compare: scripts/compare_metatomic_backends.py. Plot: scripts/plot_metatomic_backend_bench.py.

INI configuration

Fat / native

[Potential]
potential = Metatomic

[Metatomic]
model_path = /path/to/model.pt
device = cpu

Full Metatomic keys (variants, non-conservative forces, determinism) are in Metatomic Interface.

RGPOT / engine

[Potential]
potential = RGPOT

[RgpotPot]
backend = metatomic
engine_path = /path/to/libmetatomic_engine.so
model_path = /path/to/model.pt
device = cpu

Or set RGPOT_METATOMIC_ENGINE / METATOMIC_ENGINE and put the model under [Metatomic] model_path (also accepted when backend is metatomic).

Build

# Fat host: native Metatomic pot + engine for optional RGPOT consumers
meson setup build-mta -Dwith_metatomic=true -Dwith_rgpot=true
meson compile -C build-mta
# -> client/libmetatomic_pot.so  (potential = Metatomic)
# -> client/libmetatomic_engine.so  (RGPOT backend=metatomic)

# Thin host: RGPOT only; no torch at link time
meson setup build-thin -Dwith_metatomic=false -Dwith_rgpot=true
meson compile -C build-thin
# engine still comes from a fat build or a packager that includes libmetatomic_engine.so

Native Metatomic stays in eOn; it is the conda-forge path.

Runtime notes

• Prefer build-tree pot plugins over any env that also installs librgpot_pot.so (e.g. cookbook .nox/.../lib). Putting a fat install prefix first on LD_LIBRARY_PATH can interpose a pre-metatomic librgpot_pot and yield unknown backend 'metatomic'.

• The engine needs torch when dlopened (LD_LIBRARY_PATH / rpath of the engine and its metatensor stack). Thin host DT_NEEDED stays free of torch.

• The engine wraps eOn MetatomicPotential through the C ABI (rgpot_mta_create / rgpot_mta_force); energies match fat potential = Metatomic on the same model (verified on PET-MAD; see figure).