v0.1.0 — experimental, independent, not a production backend.

FP64/complex128 NumPy and optional CuPy kernels for Phono3py real-to-reciprocal FC3 transformation and normal-mode contraction. No mixed precision, frequency repair, added linewidth, altered integration weights, or production monkey patch. The adapter currently targets Phono3py 3.31.1, native r0-average=True, symmetrize_fc3q=False, full or compact FC3 and dense shortest vectors. Phono3py 4.x, other phase conventions and arbitrary WTE integrations are not validated by this release.

Use Python 3.12 in a new environment. Do not install into a production environment.

python -m venv .venv
# Activate .venv for your shell, then:
python -m pip install -e ".[dev,phono3py]"
python examples/minimal.py
python -m pytest -q
python benchmarks/benchmark_contraction.py --atoms 4 --backend numpy
python benchmarks/benchmark_fourier.py --backend numpy
python scripts/make_figures.py
python scripts/release_audit.py

The example generates a tiny synthetic algebra fixture, computes Fourier and |V3|², and compares against the separately invoked upstream native-C interaction. It does not predict a material's conductivity or solve Wigner transport. reference.py supplies a small independent complex einsum oracle. No private potential, force constants, structures, HDF5, custom WTE plugin, or accounts are needed.

For an explicitly available CUDA 12 device, install the optional CuPy extra:

python -m pip install -e ".[cuda12,phono3py]"
python examples/minimal.py --backend cupy
python benchmarks/benchmark_contraction.py --backend cupy

GPU regression is opt-in: set ACCELERATOR_TEST_GPU=1 before pytest. Check that the GPU is safely available first. Tests do not claim PASS when CUDA is absent. Do not run these commands on a shared production GPU without its owner's approval. The packaging revision has also passed an installed-wheel CPU/native-C/CUDA integration check: 24 tests, zero failures and zero skips. See packaged-wheel validation. This is a fresh synthetic correctness check, not a rerun of the private performance benchmarks below. The delivered release report binds the check to the final wheel SHA256.

中文安装与使用说明：使用说明。

from phono3py_cupy_accelerator import FourierBatch
from phono3py_cupy_accelerator.phono3py_adapter import evaluate_triplet

# pp is an existing Phono3py Interaction with solved phonons and fixed FC3.
engine = FourierBatch(pp, "numpy")  # or "cupy"; explicit choice
v3_squared, timings = evaluate_triplet(pp, triplet, backend="numpy", fourier=engine)

Eigenvectors use (vertex, component, mode), with no conjugation of the three normal-coordinate factors. The component cube interleaves atom and Cartesian indices. All three native r0 legs, native nonzero mask, all_shortest factors, prephase, signed frequency cutoff and native conversion factor are retained. No zero-g screening is introduced into the contraction. Keep FC3, mappings and grid unchanged while reusing an engine; create a new engine if inputs change.

The adapter returns per-triplet interactions, not a drop-in replacement for Phono3py's complete gamma/WTE driver. Integration weights, Bose factors, degeneracy handling, q weights, velocity operators and thickness normalization remain the responsibility of an independently validated downstream driver. No custom WTE plugin is distributed or modified here. No backend is registered globally. Default packed-constant budget is 2 GiB (not a total GPU memory bound); the phase cache is capped at 256 entries. The cubic dense algorithm remains memory-intensive and is not a sparse or large-cell solution.

The following are historical software measurements on private fixtures, not new synthetic benchmark results. Raw scientific inputs cannot be redistributed. RTX 5080 was used. Replotting the supplied summaries is reproducible; independently reproducing those private material benchmarks is not possible from this repository.

The CPU host was not exclusive. C had one warm measurement; NumPy/CuPy values are warm medians. The roughly 1098× C/CuPy ratio includes algorithmic reorganization, not pure GPU acceleration. Same-algorithm NumPy/CuPy ratio is about 2.58× for this particular microbenchmark. CuPy's first cold contraction was 0.755 s. These block-32 measurements predate the full-first-GEMM fast path packaged here; the public synthetic benchmark measures the current path instead.

The Fourier sample has 21.7× steady-state acceleration across three sampled triplets: native sum 3.628 s versus warm CuPy sum 0.1673 s. Preparation was 2.017 s and the first CuPy call was 1.766 s, excluded from that ratio. Observed GPU process peak was about 2.98 GB, not a guaranteed allocation maximum.

The historical runs were not an isolated, matched full-pipeline performance experiment. Therefore no fair end-to-end speedup ratio is claimed from 42.34/9.25 or 42.34/9.89. CuPy pool reservation and sampled process memory are different metrics; benchmark scripts label the scope instead of calling either an exact hardware peak.

The frozen ancestor completed a 73-atom / 219-branch / 24-irreducible-q M14 comparison against complete streamed NumPy using the original CPU Fourier. The refactored release has separate synthetic/native-C CPU and installed-wheel CUDA tests; it has not rerun the excluded private full-M14 inputs during packaging.

All-q gamma comparisons used rtol 1e-8 plus per-q atol equal to 1e-12 times reference peak. The modal κ test uses rtol 1e-7 and atol 1e-12 in unchanged native output units; ratio = max(|error| / (atol + rtol |reference|)). Near-zero errors are retained. Complete anonymized per-q error summaries are in benchmarks/data/gamma_accuracy_24q.csv; no linewidth arrays are distributed. The historical modal maximum absolute differences were 1.28119e-10 (κP) and 4.26048e-15 (κC) in native output units, not republished material conductivities.

The private strict-periodic 8 → 72 atom representation retained 216 branches, with the same physical sampling (8-atom M42 ↔ 72-atom M14). Harmonic and reciprocal tensor folding prechecks passed. The complete WTE comparison is accepted by the project owner at the following measured precision:

The maximum reported difference is 0.349325%. The earlier strict regression criterion (rtol 1e-7, atol 1e-12) was not met; that original result is retained alongside the owner's subsequent acceptance of these differences. The completed comparison is 72-atom M14; no 72-atom M21 result is included. Absolute material κ values and unpublished force constants are excluded.

• Public synthetic example, NumPy oracle and native-C regression: runnable.
• Public timing scripts: runnable on your explicitly available CPU/GPU; no hidden assets.
• Three figures: regenerated from the sanitized JSON/CSV, PNG + editable SVG + PDF.
• Historical private full-M14 gamma/WTE measurements: summaries only, not independently reproducible here.
• Full native-C all-grid reference: pending; no production migration approval.
• Packaging changes: removed 73-atom/4-image hardcoding, added input/version/memory guards and bounded cache; arithmetic remains the frozen three-leg/three-GEMM path.

See validation, benchmarking, release checklist, and third-party attribution. requirements/tested-cpu.txt records the isolated local validation environment.

Project author: yangdewei — 1147398767@qq.com.

BSD-3-Clause; Phono3py/Phonopy/Atsushi Togo notices are retained separately. See NOTICE and tagged upstream license. This project is not affiliated with or endorsed by upstream developers. Please cite Phono3py according to its own documentation when using it for science.

Figure QA was assisted by procedural scientific-visualization guidance described by Kassis, Agarwal, He, Patel & Brueckner (2026), Scientific Agent Skills. This procedural acknowledgment is not numerical validation or upstream endorsement.