[niklashoelter]

Single-structure UMA runs (optimize, smiles, or optimize_single_structure without a calculator) now load the Fairchem calculator with fairchem's default inference settings. Sequential runs over many structures keep "batch".

Why

Benchmark on catalyst (Quadro RTX 6000), d4rev, single ASE optimizations:

Run	"batch"	default settings
ibuprofen, first	54 s	43 s
ibuprofen, same molecule again	39 s	10.8 s
atorvastatin	202–219 s	159–169 s
12 different molecules in a row (sequential)	27 s	56–61 s

Energies agree within 0.1 meV. The default settings are faster while the composition stays the same, but they drop their compiled model at every new composition. That is why sequential runs keep "batch".

Changes

• load_calculator(config, single_structure=False): single_structure=True selects fairchem's default settings for Fairchem and is ignored by ORB and SevenNet.
• optimize_single_structure requests the single-structure calculator when it loads one itself. _optimize_sequential keeps the shared "batch" calculator.
• The calculator cache keys on the flag for Fairchem only, so ORB and SevenNet still load one model.

Testing

• Full suite: 776 passed, 1 skipped.
• New tests cover the loader settings for both loaders and both modes, which calculator each path requests, and that ORB loads once.
• gpuma optimize --smiles <ibuprofen> with the default d4rev model on cuda:3 returned ok and converged: true. fairchem logs that it runs in a reduced-precision (TF32) mode; on these Turing GPUs that has no numerical effect.