Execution Modes and Capacity

streamed_edges="direct" and execution_profile="capacity" are the default for Kokkos evaluation. The model precision defaults to FP32. Two-interaction direct execution requires an artifact matching that precision and fails closed rather than silently choosing a different algorithm. Admitted single-layer models have no R1 stage and use built-in direct execution. Their capacity plans may still compile or load specialized M0/R0 operator modules.

Prepare a direct host artifact before a reproducible CPU deployment:

export SYMMETRIX_JIT_CACHE=/path/to/private/jit-cache
symmetrix_prepare_jit_host_artifact \
    --model srtio3-mace.json --precision float32

For explicit dtype="float64" evaluation, prepare the corresponding artifact with --precision float64.

Use streamed_edges="non-compiled" only for compiler-free debugging or compatibility checks. It is not optimized, has no performance guarantee, and is not the primary execution path.

capacity ranks qualified internal MH-0 plans using an advisory device-memory estimate. It chooses the fastest estimated fit; when no candidate fits, it attempts the smallest qualified plan and lets the allocator determine final feasibility. execution_profile="speed" fixes the backend-qualified throughput plan without considering available memory. For standard compact MACE, that plan retains harmonic gradients, reuses eligible MH-0 state, and recomputes readout state. For larger graphs, CPU and CUDA FP32 select Y-only harmonics when that qualified path is faster. FP32 also uses compact unit-direction geometry. FP64 retains Cartesian geometry; CUDA FP64 and HIP retain harmonic gradients. Bounded fixed-workspace plans require the separate allow_fixed_workspace=True opt-in, currently require CUDA, and support FP32 and FP64 ordinary MH0 standard MACE. MACEField and MACE-MH-1 do not support fixed-workspace plans. They are excluded by default. In LAMMPS MPI, the ordinary two-layer standard-MACE plan uses the existing forward-H1 and reverse-H1-adjoint exchanges between its three tiled phases. Its receiver and edge workspace is bounded, while the hidden state remains proportional to the local-plus-ghost feature count. Two-layer execution requires a precision-matched generated device artifact with tiled R1 support.

from ase.spacegroup import crystal
from symmetrix import Symmetrix

calc = Symmetrix(
    "srtio3-mace.json",
    execution_profile="capacity",
    allow_fixed_workspace=True,
)
a = 3.905
atoms = crystal(
    symbols=["Sr", "Ti", "O"],
    basis=[(0, 0, 0), (0.5, 0.5, 0.5), (0.5, 0.5, 0)],
    spacegroup=221,
    cellpar=[a, a, a, 90, 90, 90],
)
atoms.calc = calc
atoms.get_forces()  # prepares the graph
print(calc.execution_plan)  # MH-0 graph-time report

low_memory=True and low_memory=False remain compatibility spellings for capacity and speed. generic is a deprecated alias for non-compiled; factorized and direct_streamed alias direct and change an otherwise-capacity request to the old speed behavior. Use literal direct for capacity behavior. materialized is the original Symmetrix execution path used by the original Symmetrix model format (named v1 here). The public mode and backend matrix is in Execution Support Matrix; implementation architecture is described in Direct Execution Architecture.

The graph-time execution-plan report currently covers MH-0 MACE and MACEField. MH-1 direct execution has its own qualified policies but does not yet expose the corresponding planner report.