v0.4.1

Primarily a bug fix release.

Release from commit 799c9fd

Python Wheels

Package	Platform	Python	ABI
moreau	any	3.9+	py3-none-any
moreau-cpu	Linux x86_64	3.9+	abi3-py39
moreau-cpu	Linux aarch64	3.9+	abi3-py39
moreau-cpu	macOS ARM64	3.9+	abi3-py39
moreau-cuda12	Linux x86_64	3.12+	abi3-py312
moreau-cuda12	Linux aarch64	3.12+	abi3-py312
moreau-cuda13	Linux x86_64	3.12+	abi3-py312
moreau-cuda13	Linux aarch64	3.12+	abi3-py312

# CPU only
pip install moreau-*.whl moreau_cpu-*.whl

# GPU (CUDA 12.2+)
pip install moreau-*.whl moreau_cpu-*.whl moreau_cuda12-*.whl

Linux aarch64 CUDA wheels target Grace Hopper (sm_90), Orin
(sm_87), and Blackwell sm_120 on CUDA 13. They are build-validated
only; run python -m moreau check after install to verify.

C/C++ source builds

Build the C API with your local toolchain from a source checkout:
packages/moreau-c/build.sh cpu or packages/moreau-c/build.sh cuda.
The output in dist/c-api/ contains the library, header, and licenses.

CUDA artifacts are built and validated on CPU runners. CUDA runtime
testing is opt-in while GPU runners are unavailable; the release QA
summary records whether it ran.

v0.4.0

ChatGPT generated:

Moreau 0.4.0 adds Direct Conic Constraints and expands native artifact distribution for Python, C, and Julia.

• Declare constraints directly on primal variables with DirectConeSpec and Cones.dir_cones, including batched CPU/CUDA solves and differentiation. See the Direct Conic Constraint guide.
• Invalid cone fields and nonfinite generalized-power weights now raise validation errors.
• CUDA solves work across the native boundary under torch.compile, including forward and backward passes.
• Python wheels cover Linux x86-64/ARM64 and macOS ARM64; C libraries and a Julia source bundle include an artifact manifest for automatic native-library resolution.
• CUDA builds retain cuDSS 0.7.1 following the measured performance and repeatability regressions in 0.8.

Install the CPU package with pip install moreau==0.4.0, or use pip install 'moreau[cuda12]==0.4.0' / pip install 'moreau[cuda13]==0.4.0' for CUDA (Python 3.12+).

See PR #2 for the implementation and detailed validation.

Artifact QA passed: Test Release v0.4.0. Hosted QA passed on Linux x86-64/ARM64 and macOS ARM64, including Python ABI checks and 2,895 Julia package assertions plus 60 integration assertions. The macOS job passed on an unchanged rerun after an isolated sub-millisecond timing comparison failed. The exact released Linux x86-64 CUDA 12 and CUDA 13 wheels each passed 180 focused local tests, with 3 existing skips, on an NVIDIA T500 using cuDSS 0.7.1.4. These tests cover cuDSS settings, CPU/GPU parity, Direct Conic Constraints, Torch compilation/gradients, and JAX. The larger pre-release GPU audit is documented in PR #2; its batch-128 backward stress case exceeded this GPU's 4 GB memory.

Artifacts from commit 36ea4ea

Python Wheels

Package	Platform	Python	ABI
moreau	any	3.9+	py3-none-any
moreau-cpu	Linux x86_64	3.9+	abi3-py39
moreau-cpu	Linux aarch64	3.9+	abi3-py39
moreau-cpu	macOS ARM64	3.9+	abi3-py39
moreau-cuda12	Linux x86_64	3.12+	abi3-py312
moreau-cuda12	Linux aarch64	3.12+	abi3-py312
moreau-cuda13	Linux x86_64	3.12+	abi3-py312
moreau-cuda13	Linux aarch64	3.12+	abi3-py312

# CPU only
pip install moreau-*.whl moreau_cpu-*.whl

# GPU (CUDA 12.2+)
pip install moreau-*.whl moreau_cpu-*.whl moreau_cuda12-*.whl

Linux aarch64 CUDA wheels target Grace Hopper (sm_90), Orin
(sm_87), and Blackwell sm_120 on CUDA 13. They are build-validated
only; run python -m moreau check after install to verify.

C API Libraries (tarballs)

Tarball	Platform	Backend
moreau-cpu-linux-x86_64.tar.gz	Linux x86_64	CPU
moreau-cpu-linux-aarch64.tar.gz	Linux aarch64	CPU
moreau-cpu-macos-arm64.tar.gz	macOS ARM64	CPU
moreau-cuda12-linux-x86_64.tar.gz	Linux x86_64	CUDA 12
moreau-cuda12-linux-aarch64.tar.gz	Linux aarch64	CUDA 12
moreau-cuda13-linux-x86_64.tar.gz	Linux x86_64	CUDA 13
moreau-cuda13-linux-aarch64.tar.gz	Linux aarch64	CUDA 13

Each tarball contains the shared library, include/moreau.h, LICENSE, and NOTICE.

CUDA artifacts are built and validated on CPU runners. CUDA runtime
testing is opt-in while GPU runners are unavailable; the release QA
summary records whether it ran.

v0.4.0-beta.1

Release from commit 20cdc2d

Python Wheels

Package	Platform	Python	ABI
moreau	any	3.9+	py3-none-any
moreau-cpu	Linux x86_64	3.9+	abi3-py39
moreau-cpu	Linux aarch64	3.9+	abi3-py39
moreau-cpu	macOS ARM64	3.9+	abi3-py39
moreau-cuda12	Linux x86_64	3.12+	abi3-py312
moreau-cuda12	Linux aarch64	3.12+	abi3-py312
moreau-cuda13	Linux x86_64	3.12+	abi3-py312
moreau-cuda13	Linux aarch64	3.12+	abi3-py312

# CPU only
pip install moreau-*.whl moreau_cpu-*.whl

# GPU (CUDA 12.2+)
pip install moreau-*.whl moreau_cpu-*.whl moreau_cuda12-*.whl

Linux aarch64 CUDA wheels target Grace Hopper (sm_90), Orin/Thor
(sm_87), and Blackwell sm_120 on CUDA 13. They are build-validated
only; run python -m moreau check after install to verify.

C API Libraries (tarballs)

Tarball	Platform	Backend
moreau-cpu-linux-x86_64.tar.gz	Linux x86_64	CPU
moreau-cpu-linux-aarch64.tar.gz	Linux aarch64	CPU
moreau-cpu-macos-arm64.tar.gz	macOS ARM64	CPU
moreau-cuda12-linux-x86_64.tar.gz	Linux x86_64	CUDA 12
moreau-cuda12-linux-aarch64.tar.gz	Linux aarch64	CUDA 12
moreau-cuda13-linux-x86_64.tar.gz	Linux x86_64	CUDA 13
moreau-cuda13-linux-aarch64.tar.gz	Linux aarch64	CUDA 13

Each tarball contains lib/libmoreau_cpu.{so,dylib} or lib/libmoreau_cuda.so.