Auto3D 3.0.0

First release published to PyPI since 2.3.1. The v3.0.0 and v3.5.0 git tags that existed previously were never published to any index; if you installed with pip or conda, every change below applies to you.

Full details in CHANGELOG.md; upgrade path in the migration guide.

Results that change

• Thermochemistry now uses most-abundant-isotope masses, matching Gaussian and ORCA. Every reported H_hartree, S_hartree_per_K and G_hartree changes. Mass enters the moments of inertia, the mass-weighted Hessian (so every frequency, ZPE and S_vib) and the translational term. Effect grows with heavy-halogen content. Cached or published thermochemistry from an earlier version is no longer bit-comparable — regenerate.
• E_tot is Hartree in every file Auto3D writes. It previously meant eV or Hartree depending on which component wrote the file, while consumers hard-coded eV.
• opt_steps below 10 is refused at construction. It was returning an unconverged structure labelled as optimized: the optimizer only tests convergence on istep % 10 == 0 and guards its statistics with n >= 10.

Fixed

• A custom NNP returning float64 forces crashed the FIRE loop — but only when two or more molecules reduced in the same step, so it was never reliably reproducible.
• smiles2mols built torch.device("cuda:N") with no bounds check, so an out-of-range GPU index produced cuda:99 instead of an error.
• reorder_sdf staged through a predictable temp name and dropped the target's file mode, so a 0600 input came back 0644 — a permission loosening.
• select_tautomers truncated its derived output path with no gate.
• Blank lines in a .smi file raised a bare IndexError or tuple-unpack ValueError.
• A malformed config file exited 1 through auto3d <config>.yaml but 2 through auto3d run -c, so a script gating on the exit code got different answers from the two entry points.
• Chunk sizing initialized a full CUDA context in the orchestrator process before workers were spawned.
• The OOM retry ran empty_cache() while the exception was still live — so it could only release already-free blocks — and then reused the original unshrunk batch size for later slices.

Performance

• import Auto3D is 44x cheaper: 1.35 s → 0.031 s, 1175 modules → 154. It no longer imports torch or RDKit. Three eager optional-dependency probes were defeating the lazy-import mechanism that exists to prevent exactly this.

• The optimizer's host↔device synchronizations drop from 18 per step to 2, and ANI2xt's from 22 per forward to 2. Proven bit-identical against a reference implementation of the old loop across 17 scenarios. The ANI2xt element loop now compiles to one torch.compile graph instead of zero (it was skipping the whole frame on a data-dependent branch).

  No wall-clock speedup is claimed. Every count above is CI-enforced on CPU; durations require a GPU and have not been measured. benchmarks/run_perf_ab.sh is provided for that. Documentation that previously claimed a ~1.25x torch.compile gain has been corrected — that path was compiling to zero graphs, so the figure could not have come from it.

Breaking API changes

• from Auto3D.utils import <name> no longer resolves for 41 names; utils/chemistry.py and utils/file_ops.py are split up. The CHANGELOG carries the full old→new mapping.
• optimizing takes a built adapter, keyword-only, and no longer resolves a model name itself. Construct the adapter inside the worker process.
• One model contract: ModelAdapter lives in Auto3D.models.contract. CustomNNP — the public custom-NNP contract — is unchanged.
• check_valid_configuration takes an Auto3DOptions instead of ten keyword arguments.
• filter_unique is deleted; the two conformer filters are one. Filters return a FilterResult that records why conformers were dropped, so "No structure converged" is no longer reported when the cause was stereochemistry.
• smiles2smi, encode_ids, decode_ids and select_tautomers take a keyword-only overwrite.
• Auto3D.cli and Auto3D.models no longer re-export names; cli/__init__.py was additionally shadowing its own app and console submodules.

Tests

1273 → 1646, with no expected failures. Four tests that could not fail were found and fixed, one of which had been concealing the exit-code divergence above by re-implementing inline the very function it claimed to exercise.