Licences

FastMACE

FastMACE is distributed under the MIT licence (LICENSE; SPEC §14, owner decision 4). The few third-party files shipped inside the package keep their own licences; they are listed with origin and licence in NOTICE.md at the top of the repository:

• the vendored OpenEquivariance receiver-streaming kernels (fastmace/kernels/oeq_streaming/, BSD-3-Clause, Lawrence Berkeley National Laboratory / University of California);
• the D3 reference tables fastmace/longrange/data/dftd3_params.npz (from torch-dftd 0.5.1, MIT, Preferred Networks); see the assessment below.

Runtime dependencies

FastMACE depends on these packages; it does not copy them. Each keeps its own licence, which applies when you install it. Values are read from the installed package metadata (METADATA, licence files) of the qualified stack in the local WSL environment ~/environments/mace_cueq_oeq on 25 October 2026; simple-dftd3 and tad-dftd3 from ~/environments/mace_cueq_latest (they are not in the hybrid environment).

Package (version checked)	Needed for	Licence (from metadata)
torch 2.14.1+cu130	base	License-Expression: Apache-2.0 AND Apache-2.0 WITH LLVM-exception AND BSD-2-Clause AND BSD-3-Clause AND BSL-1.0 AND MIT; PyTorch's own code is BSD-3-Clause style, the others are bundled third-party parts. The CUDA wheels pull in NVIDIA runtime libraries under NVIDIA's licences.
mace-torch 0.3.16	base	MIT (classifier and LICENSE.md, © 2022 ACEsuit/mace)
e3nn 0.4.4 (through MACE)	base	MIT
ase 3.27.0	base	LGPL-2.1-or-later
numpy 2.2.6	base	BSD (classifier, LICENSE.txt)
cuequivariance 0.12.0, cuequivariance-torch 0.12.0	base	Apache-2.0 (licence file; the License: field is empty)
cuequivariance-ops-cu13 / cuequivariance-ops-torch-cu13 0.12.0	fused cuEq kernels, chosen by the user's CUDA build	NVIDIA proprietary: "LICENSE AGREEMENT FOR NVIDIA MATH LIBRARIES SOFTWARE DEVELOPMENT KITS" (Other/Proprietary License). FastMACE depends on them and never vendors or redistributes them.
triton 3.8.0 (through torch)	Triton kernels	MIT
filelock 3.32.3, PyYAML 6.0.3	base / input files	filelock: MIT; PyYAML: MIT
tomli (Python 3.10 only)	TOML input files	not installed in mace_cueq_oeq, so not verified here (owner/6-T02 follow-up)
openequivariance 0.7.0	[oeq]	BSD-3-Clause
jinja2 3.1.6, cuda-bindings 13.3.1	streaming kernels (template, NVRTC)	jinja2: BSD; cuda-bindings: LicenseRef-NVIDIA-SOFTWARE-LICENSE (NVIDIA)
graph-longrange 0.4.0	[polar]	MIT (LICENSE.md; the metadata has no licence field)
torch-dftd 0.5.1	[d3], [d3-reference]	MIT (LICENSE, © 2021 Preferred Networks, Inc.)
dftd3 1.6.0 (simple-dftd3)	[d3-reference] only	LGPL-3.0-or-later
tad-dftd3 0.7.0	test reference only, not an extra	Apache-2.0

None of the LGPL packages (ASE, simple-dftd3) is copied into FastMACE; they are imported as separate installed libraries, which the LGPL permits for an MIT program.

Model weights

FastMACE never ships or redistributes model weights. fm.load downloads a checkpoint from its official source (or uses a local copy), checks its SHA-256 and shows its licence once per process as a ModelLicenceWarning (fastmace/model/labels.py, fastmace/model/loading.py). Several MACE foundation models, MACE-MH-1 among them, are under the Academic Software License (non-commercial use only); the label table records each one. The weights' licence binds you independently of FastMACE's MIT licence.

Assessment: may the release ship dftd3_params.npz?

Conductor's assessment for the owner (task 6-T02); not legal advice, and the owner decides.

What the file is. dftd3_params.npz is byte-identical to torch_dftd/nn/params/dftd3_params.npz of torch-dftd 0.5.1 (SHA-256 in NOTICE.md; tests/cpu/test_d3_kernel.py checks it). It holds numbers only: the D3 C6 reference values with their reference coordination numbers, the cutoff radii R0, covalent radii and r2/r4 by atomic number.

torch-dftd's licence. torch-dftd is MIT (© 2021 Preferred Networks), and the file is part of its released package. MIT allows copying and redistribution, also under another licence, provided the copyright notice and permission notice are kept. FastMACE keeps both: fastmace/longrange/data/LICENSE (torch-dftd's text) and NOTICE.md files, all included in the wheel through pyproject.toml's package data. As far as torch-dftd's licence goes, shipping it is permitted.

Provenance of the numbers. The values are S. Grimme's DFT-D3 reference data (Grimme, Antony, Ehrlich, Krieg, J. Chem. Phys. 132, 154104 (2010); BJ damping: J. Comput. Chem. 32, 1456 (2011)). They originate in the original program's coefficient file (pars.f, distributed with DFT-D3 by the Grimme group, Universität Bonn); exactly how torch-dftd derived its file was not traced (uncertain). Public sources show:

• The original DFT-D3 program is distributed under the GNU GPL, version 1 or later; the repackaged dftd3-lib states exactly that, and conda's packaging of the original lists it as GPL-1.0+.
• simple-dftd3 (the Grimme group's current implementation, which carries the same reference data) is LGPL-3.0-or-later.
• We found no separate licence placed on the reference numbers as a dataset.

Open questions for the owner.

1. Whether a table of physical reference values (computed C6 coefficients) is copyrightable at all is jurisdiction-dependent. In the EU a database right may also apply. If it is protected, the upstream licence of the data is the GPL (original program) or LGPL (simple-dftd3), and torch-dftd's MIT relicensing of the converted file would only be valid if those terms allowed it. They would not, for a copyleft licence. That is torch-dftd's question, not one FastMACE can settle.
2. Low-risk options, for the owner to choose: (a) ship as now, with torch-dftd's MIT notice plus an attribution to Grimme et al. (done in the data NOTICE.md and here), relying on torch-dftd's licence, as many MIT/BSD D3 implementations (torch-dftd itself, and others built on the same tables) do; (b) not ship the file and fetch it at first use from the installed torch-dftd (making [d3] a requirement of the native D3 path, against SPEC §4.1); (c) ask the Grimme group (or Preferred Networks) for a statement about the data's licence before 1.0.
3. Ported ASE code: the FastMACE integrators and optimisers port ASE 3.27's algorithms to the GPU for agreement to rounding. If any port is a close enough translation to be a derived work of LGPL-2.1-or-later code, that file would be LGPL, not MIT. Decided (owner D8.2, 25 October 2026; task 6-T02b): the ports are marked LGPL-2.1-or-later. The 13 files are: fastmace/integrators/{langevin,baoab,berendsen,csvr,nose_hoover,mtk,nve}.py, fastmace/optimize/{fire,lbfgs,cell}.py, fastmace/reference/integrators.py and fastmace/kernels/triton/{fused_integrator,nose_hoover_chain}.py. Each carries # SPDX-License-Identifier: LGPL-2.1-or-later and a line naming the ASE source it ports; NOTICE.md lists them, the licence text is LICENSES/LGPL-2.1-or-later.txt, and pyproject.toml declares license = "MIT AND LGPL-2.1-or-later". tests/cpu/test_lgpl_headers.py keeps the headers, NOTICE.md and the licence expression in step. Found by searching fastmace/ for ASE provenance in docstrings ("ASE 3.27's ...", "port"); the fastmace/ase/ shims mirror ASE's signatures and stay MIT.

No change to what ships was made by this task.