dendra-solvers

C++ PyTorch extensions for fast, differentiable implicit solvers in Dendra.

Installation

pip install dendra-solvers

Python 3.10+ and PyTorch 2.12+ are required. When a compatible wheel is available, installation does not require a compiler. Stable-ABI wheels work across supported Python and PyTorch versions on the same operating system and CPU architecture.

For an HPC environment or custom PyTorch build, install PyTorch first and build the extension against that installation:

python -m pip install "torch>=2.12.0" "setuptools>=70.1,<82" ninja wheel
MAX_JOBS=1 python -m pip install --no-build-isolation --no-cache-dir \
  --no-deps --no-binary=dendra-solvers dendra-solvers

The build uses the C++ standard selected by your installed PyTorch. Use a C++20-capable compiler for current PyTorch releases. Stable-ABI builds do not need rebuilding just because PyTorch is upgraded; see installation and clean rebuilds for wheel availability, source builds, and upgrading older installations.

Useful build options:

• DENDRA_NATIVE=1 enables host-specific compiler optimizations. Do not use it for redistributable builds.
• DENDRA_OMP=1 enables OpenMP when its runtime is available.
• MAX_JOBS=1 limits compilation concurrency on memory-constrained systems.

Testing

For a development checkout:

python -m pip install "torch>=2.12.0" "setuptools>=70.1,<82" ninja wheel numpy pytest
MAX_JOBS=1 python -m pip install -v --no-build-isolation --no-deps -e .
python -c "import dendra_solvers._ext as ext; print(ext.__file__)"
python -m pytest -q -ra

Solver contracts and PyTorch transform usage are documented in docs/.

See Contributing for development tools, commit conventions, and versioning.