Holoscan Vision for CUDA 13

This distribution provides holoscan_vision and its native Vision library. It requires the exact matching holoscan-core-cu13 wheel, cvcuda-cu13 for CV-CUDA, and nvidia-npp>=13.0,<13.2 for NPP. These are declared dependencies, so pip installs them with it. Other CUDA 13 runtime prerequisites still apply, and the CUDA driver is supplied by the host and is never bundled.

The bindings expose schema-derived Python proxies for all eight CV-CUDA-backed operators and the NPP-backed BayerDemosaicOp. Import them from holoscan_vision.operators; the Bayer proxy accepts stable phase names such as bayer_grid_pos="rggb" and carries holoscan.schema.Image on both ports.

Build from a completed SDK build tree using regular CPython 3.12:

export CUDA_MAJOR=13 PEP440_VERSION=5.0.0.dev1
python3.12 -m build --wheel python/holoscan-vision \
  -Ccmake.define.HOLOSCAN_SDK_BUILD_ROOT=/absolute/path/to/sdk/build

The shared scikit-build-core project installs the native build's Vision wheel component without reconfiguring or compiling the SDK or regenerating stubs. Source archives contain the staging tooling and licenses, but still require the completed native build tree and the paths referenced by its install rules.

Install Core and Vision together from the same release. Core owns holoscan/; Vision owns holoscan_vision/, including lib/libholoscan_vision.so.5. Core libraries are not duplicated. CV-CUDA and NPP resolve from their provider wheels installed alongside Vision: cvcuda_cu13.libs/ and nvidia/cu13/lib/, respectively. NPP retains its CUDA 13 SONAMEs. Staging gives Vision an origin-relative RPATH and a direct dependency on NPP's core library so both NPP libraries resolve from the provider wheel ahead of a system CUDA directory in LD_LIBRARY_PATH. See the runtime ownership policy. Both wheels target cp312-abi3 and are qualified on regular CPython 3.12–3.14. Uninstalling Vision leaves Core's files intact.