Optional Dependency Matrix¶
The base package must stay usable without heavyweight accelerator, vendor, or interop stacks. Optional dependencies are grouped by the workflow that needs them. This matrix records the declared install profile, the import checked by tests, and the current CI/dev boundary.
| Dependency | Import surface | Install profile | CI/dev boundary | Evidence |
|---|---|---|---|---|
gdsfactory>=9.0 |
gdsfactory |
pip install "sc-neurocore[optics]" |
Dedicated CI job test-optics-extra installs .[dev,optics] and runs tests/test_optics -q -rs so GDSII round-trip tests execute with gdsfactory; default CPU jobs still skip GDSII tests when the extra is absent. |
tests/test_optics/test_gdsii.py, tests/test_optics/test_photonic_emitter_branches.py, .github/workflows/ci.yml |
dwave-neal + dimod |
neal, dimod |
pip install "sc-neurocore[annealing]" |
Dedicated CI matrix lane test-optional-extras (annealing) installs .[dev,annealing] and runs tests/test_bridges/test_quantum_annealing_neal_parity.py -q -rs; default CPU jobs still skip the parity selector when either package is absent. |
tests/test_bridges/test_quantum_annealing_neal_parity.py, .github/workflows/ci.yml |
onnx |
onnx |
dev, full, research |
Dedicated CI matrix lane test-optional-extras (onnx-protobuf) installs .[dev] and runs tests/test_export/test_onnx_exporter_protobuf_graph.py -q -rs; JSON export tests run without ONNX, while protobuf export tests skip when ONNX is absent outside that lane. |
tests/test_export/test_onnx_exporter_protobuf_graph.py, tests/test_export/test_onnx_export.py, .github/workflows/ci.yml |
lava-nc |
lava on Python <3.11 |
pip install "sc-neurocore[lava]" on supported Python |
Loihi/Lava adapter packages are handoff artefacts; the dependency is recorded in generated adapter manifests rather than imported by default CI. | tests/test_nir_neuromorphic_adapters.py |
snntorch |
snntorch |
Not declared in pyproject.toml; install snntorch==0.9.4 manually for the reproduced NIR interop profile. |
Verified with nir==1.0.8 and nirtorch==2.6. snnTorch 1.0.0 currently fails inside its upstream NIR exporter before SC-NeuroCore receives a graph, so it is not a supported interop profile. The exporter test skips when snnTorch is absent outside an explicit interop run. |
tests/test_nir_bridge/test_scalar_broadcast.py, docs/guides/nir_integration.md |
spikingjelly |
spikingjelly.activation_based |
Not declared in pyproject.toml; install upstream commit 2797c5575515b59d7f09a3d5b732d6d25e148d13 manually for NIR export support. |
Verified with nir==1.0.7. NIR 1.0.8 removed input_type/output_type from nir.LIF while this upstream exporter still supplies them, so the latest/latest pairing fails before SC-NeuroCore receives a graph. The test skips when SpikingJelly is absent outside an explicit interop run. |
tests/test_nir_bridge/test_spikingjelly_interop.py, docs/guides/nir_integration.md |
cupy-cuda12x>=12.0 |
cupy, cupyx |
pip install "sc-neurocore[gpu]" |
GPU path depends on local CUDA compatibility; default and CI CPU paths must not import CuPy. PyTorch training uses its separate training extra, and auto_device() falls back when the installed PyTorch build does not support the local GPU compute capability. |
pyproject.toml, docs/guides/faq.md, docs/api/training.md |
mpi4py>=3.0 |
mpi4py |
pip install "sc-neurocore[mpi]" |
Dedicated CI matrix lane test-optional-extras (mpi) installs Open MPI plus .[dev,mpi] and runs tests/test_mpi_runner_real.py -q -rs; distributed execution outside CI requires a local mpirun runtime. |
tests/test_mpi_runner_real.py, tests/_mpi_helpers/mpi_runner_worker.py, .github/workflows/ci.yml |
Rules¶
- Do not add a package to the base dependency set just to satisfy an optional test.
- If a test uses
pytest.importorskip, the matrix must name the import and the install profile or state that the dependency is manual. - If an optional dependency becomes part of a declared profile, update
pyproject.toml, this matrix, and the focused matrix contract test together. - CI jobs that install optional profiles must name the profile explicitly and run the matching focused test selector.