Execution Support Matrix

This matrix describes the public execution algorithms. Internal kernels, storage policies, and historical selector names are not separate public modes.

Public request	Model boundary	Kokkos backends	Runtime artifact	LAMMPS
direct	Compact standard MACE and MACEField with admitted contracts; compatible compact MACE-MH-1	Serial, OpenMP, CUDA, HIP	Standard two-interaction MACE/MACEField requires an R1 artifact; single-layer standard MACE runs without R1 specialization, but capacity may specialize M0/R0; MACE-MH-1 requires its generated program	Standard MACE and MACEField; R1 artifact required for two-interaction models, but not single-layer models; MACE-MH-1 is not supported
non-compiled	Compact standard MACE and MACEField; compatible compact MACE-MH-1	Serial, OpenMP, CUDA, HIP	None	Compiler-free debug and compatibility checks only
materialized	Models in the original Symmetrix format (named v1 here) and retained numerical controls	Serial, OpenMP, CUDA, HIP	None	Original Symmetrix and supported standard MACE/MACEField controls
auto	Compatibility request resolved from model format and evaluator capability	As resolved above	As resolved above	Format-aware compatibility default

The non-Kokkos evaluator selected by use_kokkos=False does not provide prepared direct execution: a direct request resolves to serial non-compiled execution with a warning. An explicit materialized request remains materialized.

generic and all_interactions are deprecated aliases for non-compiled. factorized and direct_streamed are compatibility aliases for direct and, in the Python frontend, preserve the former throughput-oriented behavior by changing an otherwise-capacity request to the speed profile. Use the literal direct request for capacity behavior. These aliases do not emit a deprecation warning. The removed receiver_factorized selector is rejected.

Ordinary MACE exposes energy, free energy, per-atom energies, forces, and stress. MACEField additionally exposes polarization and admitted analytical Born effective charge and polarizability calculations. On prepared and generic execution the analytical response evaluates its radial splines inline from coefficient tables instead of materializing edge-sized radial caches, so response memory is dominated by node-sized adjoint tangents; SYMMETRIX_DEBUG_RESPONSE_MATERIALIZED_RADIAL=1 restores the former materialized caches for diagnosis. Property, precision, and MPI qualification can be narrower than implementation availability; dated benchmark reports provide the evidence for a specific backend and workload.

execution_profile="capacity" and "speed" select resource objectives inside the direct algorithm. allow_fixed_workspace=True separately permits bounded tiled workspace for qualified ordinary MH0 standard MACE models on CUDA in FP32 and FP64. MACEField and MACE-MH-1 do not support fixed-workspace plans. These controls do not create additional execution modes.

LAMMPS mpi_message_passing supports the dual-layer fixed-workspace plan for ordinary two-interaction standard MACE with CUDA FP32 or FP64 direct execution and a precision-matched generated tiled R1 device artifact. The receiver- and edge-sized workspace is bounded, while H1 and its adjoint remain sized to all local and ghost feature nodes. HIP, field-coupled models, parameter gradients, and execution observers are outside this qualification boundary.

See Direct Execution Architecture for the architecture, Execution Modes and Capacity for user controls, and LAMMPS Integration for the narrower LAMMPS boundary.