vibespatial Core Integration¶

vibespatial-raster depends on vibespatial for core GPU infrastructure. This document maps which core modules are used and why. See also the architecture overview and kernel developer guide.

Core modules used¶

vibespatial module	Used by	Purpose
`residency`	`buffers.py`	`Residency`, `TransferTrigger`, `select_residency_plan`
`runtime`	`buffers.py`	`RuntimeSelection`, `has_gpu_runtime`
`cuda_runtime`	`algebra.py`, `rasterize.py`	`get_cuda_runtime`, `compile_kernels`, `KERNEL_PARAM_*`
`fusion`	`polygonize.py`	`PipelineStep`, `StepKind`, `plan_fusion`

Core changes required in vibespatial¶

Two small additions to vibespatial core enable the standalone package:

__init__.py: pkgutil.extend_path(__path__, __name__) at the top, allowing Python to discover vibespatial.raster from a separate installation.
fusion.py: StepKind.RASTER = "raster" enum value for raster pipeline steps.
io_support.py: IOFormat.GEOTIFF and IOFormat.COG enum values with HYBRID path classification (optional – only needed if raster IO plans go through the core IO support matrix).

Module structure¶

vibespatial.raster/
    __init__.py          # Public API with lazy imports
    buffers.py           # OwnedRasterArray, GridSpec, ZonalSpec, PolygonizeSpec
    io.py                # read_raster, write_raster (requires rasterio)
    algebra.py           # Local ops (CuPy) + focal ops (NVRTC stencil)
    zonal.py             # Zonal statistics via segmented reduce
    rasterize.py         # Vector-to-raster (CPU + GPU paths)
    label.py             # Connected component labeling, sieve, morphology
    polygonize.py        # Raster-to-vector pipeline
    kernels/
        __init__.py
        focal.py         # NVRTC convolution kernel source