Atomli 0.1.7 — independent engines

Atomli 0.1.7 — independent engines

This release provides a complete, verified macOS Apple Silicon family of four matching wheels. Standard CPython 3.10–3.14 was tested. Linux and Windows split wheels are not included; earlier monolithic assets are not mixed with this family. Distribution is through GitHub Releases. The optional engine packages were not uploaded to PyPI.

Package split

Package	Contents	Compressed wheel size
atomli	Structures, constraints, optimizers, MD, I/O, LJ and D3	2.55 MB
atomli-tb / atomli_tb	GFN2-xTB, g-xTB and D4	6.31 MB
atomli-dft / atomli_dft	Molecular and periodic QC	23.64 MB
atomli-mlip / atomli_mlip	Machine-learning potentials	4.95 MB

Each engine has an independent build graph. The core has no QC, TB or MLIP engine dependency. Public calculator imports remain from atomli.calculators import QC, XTB, MLIP, DFTD4. Engine versions and binary protocols are checked against the core. Broken installed plugins report explicit errors; missing optional engines do not affect core workflows. Model weights remain separate runtime downloads.

Repairs included

The split boundary preserves explicit charge/spin metadata, calculator settings, retained state and cached-property semantics. Phonon force-cache identity includes the actual engine binary, and phonon work preserves independent calculator state. GPU g-xTB force evaluation now receives the actual final exchange source produced by resident SCF. Both direct-force and energy-then-force calls pass, including changed geometry, reset and CPU parity. Energy-only requests keep their existing readback contract. The release publisher also rejects mislabeled platforms, filename/metadata mismatches, altered wheels and incomplete engine families.

Verification

The final installed family passed 149 battery/split-wiring cases, 108 compatibility/settings/stub cases and 12 wheel-runtime cases. An isolated core-only environment passed 204 workflow cases, with one unsupported ASE cubic-HCP reference skipped. Release, dependency-graph and downloader contracts passed 21 cases. Runtime smoke checks passed on CPython 3.10, 3.11, 3.12, 3.13 and 3.14. CPU/GPU TB and MLIP checks include nonzero forces; the checked MLIP maximum force difference was 1.18e-6 eV/angstrom. Focused test deadlines and numerical tolerances were preserved. These checks do not establish battery-voltage accuracy or a general wall-time speedup.

Install

Download download-engines.py from this release, then run:

python -m pip install packaging
python download-engines.py --tag v0.1.7 --engines all --install

Use --engines none for core only, dft, tb, or mlip for one engine, or dft,tb for QC+D4. The downloader selects the compatible platform family and verifies every wheel against release-manifest.json. SHA256SUMS covers the wheels, manifest and downloader. Core and engine versions must match exactly.

Private Rust/WGSL sources, native build snapshots and MLIP model weights are excluded from these release assets. Free-threaded CPython, PyPy and additional target platforms require separate artifacts and verification.

Atomli 0.1.6

Atomli 0.1.6 updates molecular and periodic QC memory planning and GPU work reuse.

• RI cache admission accounts for retained topology and primitive plans.
• Response calculations stream AO blocks when retained data exceeds the working allowance.
• Geometry replacement releases the previous state before the next SCF calculation.
• Direct integral storage follows the caller's allowance.
• PBE skips unused kinetic-energy-density work.
• DIIS and molecular grids avoid duplicate storage.
• Restricted molecular QC reuses GPU state and force projections across geometry changes.

The working allowance is not a total process memory limit.
Mandatory matrices and allocator overhead can exceed it.
Large Newton response calculations recompute exact integral contractions; their large-system runtime remains unmeasured.
Numerical tolerances, basis support, and Python requirements remain unchanged.

The release provides five cp310-abi3 wheels for standard CPython 3.10 and later.
The release manifest records installed-wheel checks on macOS ARM64 and Linux x86-64.
Linux ARM64 and both Windows targets have static verification only because target runtimes were unavailable.
Private Rust source and model weights remain outside the wheels.

The attached manifests record source hashes, wheel hashes, and numerical validation results.

Atomli 0.1.5

Atomli 0.1.5 corrects Vulkan shader descriptor bindings for QC calculations.
Caffeine SCF and periodic TiO₂ calculations now pass the retained comparison fixtures on NVIDIA A100.
The macOS Metal wheel passes the same four cases on Apple M4 Max.

The comparison covers caffeine PBE, r²SCAN, SKALA 1.1, and periodic TiO₂ PBE.
It checks energy and forces across three changed geometries with retained calculator state.
TiO₂ also includes stress.
The A100 TiO₂ maximum energy error is 0.0000437 eV.
Acceptance limits remain 0.1 eV, 0.1 eV/Å, and 0.02 eV/ų.

Both native wheels pass 69 package tests on each CPython version from 3.10 through 3.14.
All five wheels pass metadata, native binary, and stable ABI checks.
Linux ARM64 and Windows x86-64/ARM64 have static verification only.
No target runtime was available for those three platforms.

Install from PyPI:

python -m pip install --upgrade atomli==0.1.5

The cp310-abi3 wheels support standard CPython 3.10 and later.
Free-threaded CPython and PyPy require separate validation.
Wheels contain compiled binaries and Python modules.
Model weights remain runtime downloads.

manifest.json records artifact hashes, runtime coverage, and working-tree source hashes.
qc-validation.json records the installed-wheel numerical results and reference hashes.
SHA256SUMS covers the wheels and evidence files.
These checks do not establish a performance improvement.

atomli 0.1.2 wheels

Atomli assets and models.

mlip-models

Atomli assets and models.