NZCVM

New Zealand Community Velocity Model: tools for building and querying tetrahedral velocity models.

A velocity model is a collection of tetrahedral meshes. Each mesh holds cell data for seismic velocity (Vp, Vs), density (rho) and the quality factors (Qp, Qs), plus a priority that controls blending where meshes overlap. nzcvm generate samples those meshes onto a structured 3D grid. It then pushes the result through a chain of layers and writes it in a format a simulation code can read.

One nzcvm generate run over a 260 × 200 km Canterbury domain. EP2020 tomography blended with fourteen basin and volcanic models, a 1-D offshore profile, the Ely Vs30 taper, a Vp/Vs clamp, and Backus averaging over each depth cell. The low-velocity wedge thickening eastward is the Canterbury Plains sediment sequence. The high-Vs lens interrupting it on the second section is the Banks Peninsula volcanics.

---

Build

The core query engine is a Rust extension built with setuptools-rust. Requires Python 3.13 or newer.

uv is the preferred build tool for this repo:

uv sync   # creates a venv and builds the Rust extension

Or build and install the wheel by hand:

pip install build
python -m build --wheel
pip install dist/*.whl --force-reinstall

The extension compiles with the release profile by default. Set SETUPTOOLS_RUST_CARGO_PROFILE=dev for debug builds.

Non-pip dependencies

Dependency	Purpose
Rust toolchain (stable)	Compiling the extension (if building from source)
HDF5 ≥ 1.12	Needed at runtime by h5py

pyproject.toml declares the Python dependencies. The base install covers nzcvm generate. Commands with heavy, single-purpose dependencies need an optional extra, and say which one when it's missing:

Extra	Installs	Needed for
vis	pyvista	nzcvm view
mesh	gmsh	nzcvm basin
raster	rioxarray (rasterio, GDAL)	nzcvm convert-tiff
bench	geopandas, joblib, platformdirs	nzcvm tree-stats

pip install 'nzcvm[vis]'     # also installs pyvista
uv sync --extra mesh         # in a checkout
uv sync --all-extras         # everything

---

Quick start

uv run nzcvm generate examples/2014p240655.toml output.zarr

The config selects the grid, the layer chain, and the models to query. The output extension selects the writer. The preceding example needs resources/ (DEM, Vs30, coastline) and models/ to be present.

With no data to hand, build the synthetic dataset instead, then run against it:

just synthetic
uv run nzcvm generate examples/synthetic.toml synthetic/model.zarr

---

Command line

Command	Purpose
nzcvm generate	Generate a velocity model from a config file
nzcvm view	Interactive 3D PyVista viewer for model output
nzcvm basin	Construct a tetrahedral mesh for a basin model
nzcvm tomography	Convert a CSV-like tomography model to a tetrahedral mesh
nzcvm surface	Convert an HDF5 topography surface to a VTK unstructured grid
nzcvm convert-tiff	Convert a GeoTIFF raster to a surface
nzcvm tree-stats	Benchmark BVH query performance
nzcvm synthetic	Write synthetic DEM, Vs30, basin, and tomography inputs
nzcvm plugins	List available layers, grids, and formats, and their packages

Useful generate options:

Option	Effect
--n-threads	Query threads (defaults to the process CPU affinity)
--format	Force an output format instead of inferring it from the path
--config-format	Force toml / yaml / json instead of inferring
--quantise	ZFP-compress the arrays in NetCDF/Zarr output
--distributed	Run on a local Dask distributed cluster
--progress	tqdm progress bar (not compatible with --distributed)
--monitor	Log CPU/memory usage while running
--log-level	DEBUG / INFO / WARNING (default WARNING)

---

Configuration

A config is a TOML, YAML or JSON file with three sections: metadata, grid, and an ordered list of layers. Config objects are plain dataclasses deserialised by mashumaro, so you can also build them in pure Python. They validate bounds, layer ordering, and layer dependencies.

Grid types

Set by grid.type. All of them are topography-following and chunked lazily with Dask.

Type	Key parameters
sw4	extent_x/y, refinements (2:1 nested resolutions, ordered automatically)
regular	extent_x/y, thickness, resolution_x/y/z (fixed vertical resolution)
emod3d	nx, ny, nz, resolution, decay
borehole	sites, depth, resolution_z (one vertical profile per site)

Every grid takes surface (path to a DEM) and optional [grid.chunks]. The three volumetric grids fill a rotated box, so they also take an [grid.orientation] block naming the model origin of that box:

[grid.orientation]
crs = 'EPSG:2193'     # target projected CRS
azimuth = 39.0        # clockwise from north, degrees
origin_lon = 176.00145
origin_lat = -39.65225

[grid.chunks]
i = 256
j = 256

A borehole grid holds independent columns with no box to orient, so it takes a bare [grid.projection] instead: just a CRS.

Terrain decay

The sw4 and emod3d grids share one terrain-following coordinate. A level at nominal depth A (its depth in a flat box) lies at

z = A + B(A, S) · S        (S = surface, positive down)

The decay weight B sets how quickly the levels stop following the topography. B = 1 follows it exactly and B = 0 is flat. Choose the decay with a [grid.decay] table:

type	Weight B	Parameters
squashed	1: the surface shifted down	none
linear	1 - A/L, flat below L	length (default: bottom of the first block)
tapered	1 - A/(ratio · elevation) on land, 1 at sea	ratio (default 1)
sleve	sinh((L - A)/s) / sinh(L/s), flat below L	scale, length (default: bottom of the first block)

An emod3d grid requires a decay. EMOD3D's squashed and squashed_tapered topography types are squashed and tapered (with the default ratio = 1).

The builder rejects a decay that folds a column over itself. This can happen with sleve under deep water: keep scale · tanh(length / scale) deeper than the deepest seafloor.

Solver coordinates

Every grid samples the earth at its true z. The grid's solver attribute says where the solver places each sample, and the sfile writer writes its interfaces in those coordinates. It follows from the decay:

• No decay (sw4 default): physical, at the true z, for SW4 with a topography command. The builder stretches the first refinement linearly between the surface and its flat bottom, so each column is evenly spaced, as SW4's sfile reader assumes.
• Any decay (and every emod3d grid): nominal, at the nominal depth A in a flat box. EMOD3D works this way, and SW4 does too without a topography command, so this is SW4's EMOD3D-compatible mode. The decay is a deliberate distortion of the earth into the box. Each block's nominal spacing is its refinement resolution.

[grid]
type = "sw4"

[grid.decay]   # nominal: run SW4 without topography
type = "tapered"

The sw4, emod3d, regular and borehole grids store their nominal depths as a nominal_depth coordinate on k, so NetCDF, Zarr, CSV, and Parquet outputs include both sets of coordinates.

Borehole grids

grid.type = "borehole" extracts one vertical profile per site rather than filling a volume. Each column runs from the topography down to depth at a fixed resolution_z. The profiles line up sample for sample and compare directly.

[grid]
type = "borehole"
surface = "./resources/dem.zarr"
depth = 600.0          # metres below the topography
resolution_z = 25.0    # metres between samples

[grid.projection]
crs = 'EPSG:2193'      # CRS the profiles are extracted in

[[grid.sites]]
longitude = 172.6218
latitude = -43.5283
site = "CACS"          # not a keyword; see below
network = "NZ"

Sites come in a global CRS (WGS84 unless sites_crs says otherwise), and the builder maps them into grid.projection.crs. Instead of listing them inline, point sites at a CSV or Parquet file with longitude and latitude columns:

sites = "stations.csv"

Keep that line ahead of [grid.projection]: TOML would otherwise read it as a key of that table.

Site labels

Longitude and latitude place a site, and the config reserves nothing else. Every other key, and every other column of a site file, becomes a coordinate on the grid's i axis under the name the caller gave it, which the writers keep. In the preceding example, site and network are labels rather than keywords.

A label keeps the type the caller wrote: the writers store a numeric column as numbers, not as text. Each site needs the same set of labels, since the alternative is a column of nulls where one site was missing a key.

A label may not take a name in nzcvm.grids.grid.RESERVED_COORDINATES, which covers everything GridSchema declares, the (i, j, k) index, and the component names. Naming one raises rather than corrupting the grid.

To drop the labels and keep only the spatial coordinates:

keep_extra_columns = false

The result has shape (len(sites), 1, nk): one column per site, with nk samples down each column. The singleton j axis preserves the (i, j, k) contract every layer relies on. The site labels index the i axis, so the output reads back per station:

import xarray as xr

tree = xr.open_datatree("boreholes.zarr", engine="zarr")
grid = tree["grids/boreholes"].ds
vs = tree["qualities/boreholes"].ds.vs.squeeze("j")

for n, site in enumerate(grid.site.values):
    print(site, vs[n].values)

Writing to a *.csv or *.parquet path instead puts those labels in the header, one row per sample:

uv run nzcvm generate examples/borehole.toml boreholes.csv

grid,site,network,i,j,k,x,y,z,depth,rho,vp,vs,qp,qs,alpha
boreholes,GULL,NZ,0,0,0,1531509.5,5161095.5,-641.124146,0,1810,1800.00012,500,100,50,1
boreholes,GULL,NZ,0,0,1,1531509.5,5161095.5,-616.124146,25,1810,1800,500,100,50,1

which pandas groups straight back into profiles:

import pandas as pd

table = pd.read_csv("boreholes.csv")  # or read_parquet("boreholes.parquet")
for site, profile in table.groupby("site", sort=False):
    print(site, profile.vs.to_numpy())

examples/borehole.toml runs four profiles over the just synthetic dataset, two of them inside a basin.

Layers

Layers run in the order listed, outermost first. Each one delegates down the chain, so the last entry must be query. Some layers declare dependencies: ely and offshore both need the coastline coordinate, so a coastline layer has to appear before them. The config raises a validation error if a dependency is missing.

Type	Description
query	Queries the tetrahedral model tree (required, and always last)
ely	Ely et al. (2010) near-surface Vs taper from a Vs30 map
offshore	1-D offshore/coastal velocity profile (requires coastline)
coastline	Computes signed distance to the coastline, provides coastline
clamp	Clamps components and the Vp/Vs ratio to physical bounds
backus	Alpha-weighted super-sampling over each depth cell
custom	Any layer registered with @functional_layer or as a Layer subclass

Output formats

Inferred from the output path, or forced with --format.

Format	Path	Description
zarr	*.zarr	Chunked array store, keeps all metadata, good for debugging
netcdf	*.h5	NetCDF4/HDF5 via xarray
sfile	*.sfile	sfile HDF5 format for driving SW4
emod3d	directory	rho3dfile.d, vp3dfile.p, vs3dfile.s binaries suitable for driving EMOD3D
csv	*.csv	Flat table, one row per point, labelled by grid, and by site
parquet	*.parquet, *.pq	The same table, with the float32 columns kept typed

Plugin packages can add more formats (see Plugins).

csv and parquet share one flattening step, so the columns are the same either way. Both hold the whole table in memory, which suits the outputs a person reads: boreholes, transects, a few profiles. Volumetric grids belong in Zarr or NetCDF.

Example configuration

See examples/ for complete, working configs:

File	Shows
2014p240655.toml	SW4 grid, full layer chain
2014p240655_emod3d.toml	The same domain as an EMOD3D grid
whole_country.toml	regular grid over New Zealand
near_fault_config.toml	A custom layer (examples/near_fault.py) in a config
synthetic.toml	The same chain over the just synthetic dataset
borehole.toml	borehole grid: profiles at four sites

[metadata]
title = "Domain for geonet event 2014p240655"

[grid]
type = "sw4"
surface = "./resources/dem.zarr"
extent_x = 210000.0
extent_y = 330000.0

[grid.orientation]
crs = 'EPSG:2193'
azimuth = 39.0
origin_lon = 176.00145
origin_lat = -39.65225

[grid.refinements.top_layer]
resolution = 200.0
bottom = 5000.0     # bottom elevation, +z down

[[layers]]
type = "clamp"
min_vp_vs_ratio = 1.73
max_vp_vs_ratio = 4.0

[[layers]]
type = "coastline"
coastline = "resources/coastline.wkb.gz"

[[layers]]
type = "ely"
vs30 = "./resources/vs30.zarr"
depth_t = 450.0

[[layers]]
type = "query"
model_path = "./models"
model_globs = ["*.zarr"]

---

Reading the output

Zarr and NetCDF output is a standard xarray DataTree with two top-level groups, each with one node per grid:

/
├── grids/<name>       x, y, z, depth   (i, j, k)
└── qualities/<name>   vp, vs, rho, qp, qs, alpha   (i, j, k)

x, y, z and depth are data variables on the logical (i, j, k) index, not dimension coordinates. The grid is curvilinear so each coordinate is a 3D array.

import matplotlib.pyplot as plt
import xarray as xr

dt = xr.open_datatree("output.zarr", engine="zarr")
grid = dt["grids/grid"].ds
qual = dt["qualities/grid"].ds

# Horizontal Vs slice 200 m down (k = 2 at 100 m vertical spacing)
plt.pcolormesh(grid.x[:, :, 2], grid.y[:, :, 2], qual.vs[:, :, 2])

# West–east vertical cross-section at j = 100
plt.pcolormesh(grid.x[:, 100, :], -grid.z[:, 100, :], qual.vs[:, 100, :])

This block generates a depth slice and a cross-section through an 80 × 80 × 20 km Wellington domain at 400 m horizontal resolution. Both sit high in the section because the Wellington basins are shallow: the low-velocity cap runs 100 to 400 m thick over most of the domain, so a slice at 500 m cuts almost entirely below it.

For interactive 3D visualisation with PyVista (requires nzcvm[vis]):

nzcvm view model output.zarr --scalar vs --coastline resources/coastline.wkb.gz

nzcvm view model also diffs two models (--compare-to, --diff-mode) and can render off-screen to an image file (--off-screen --screenshot out.png).

---

Python API

Query a model tree directly, without building a grid:

from pathlib import Path
from nzcvm.models.model import ModelTree

tree = ModelTree.load_models(
    [Path("models/ep2020.zarr"), Path("models/Wellington.zarr")]
)
quality = tree.query(x=1_749_150.0, y=5_428_150.0, z=500.0)
print(quality.vp, quality.vs)  # None if the point is outside every mesh

load_models takes an iterable of mesh paths, anything xarray can open (the meshes distributed in models/ are Zarr). Coordinates are in the model's projected CRS with z positive downwards.

query_many is the vectorised form and returns a Qualities dataset. explain shows which models contributed to the blend:

>>> tree.explain(1_749_150.0, 5_428_150.0, 100.0)
(ρ=1810.00, Vp=1800.00, Vs=580.00, Qp=58.00, Qs=29.00, ɑ=1.00)
├── Model 0 (priority = 38)
│   └── Quality: (ρ=1810.00, Vp=1800.00, Vs=580.00, ...)
└── Model 1 (priority = 255)
    └── Quality: (ρ=2539.03, Vp=4564.84, Vs=2630.92, ...)

Lower priority numbers win. Overlapping models are alpha-composited until the cumulative alpha reaches 1.0. ModelRange restricts a query by priority band: BASINS is 0-127, TOMOGRAPHY is 128-255, ALL is both.

---

Testing

uv run pytest tests/
uv run pytest --doctest-modules nzcvm/   # doctests
uv run ruff check nzcvm/ tests/          # linting
uv run ty check nzcvm/                   # type checking

Tests marked real_data need a model directory supplied via MODEL_PATH.

Synthetic data

just synthetic writes a self-contained data root under synthetic/, so a run of nzcvm generate covers the whole path without the real NZCVM data. It takes about a minute, most of it meshing the basins, and needs no NZCVM_DATA_ROOT.

Recipe	Writes
synthetic_dem	dem.h5, dem.zarr (topography for grid.surface)
synthetic_vs30	vs30.h5, vs30.zarr (Vs30 map for the ely layer)
synthetic_coastline	coastline.wkb.gz (land polygon in NZTM)
synthetic_tomography	models/tomography.zarr (background tomography mesh)
synthetic_basins	models/gully.zarr, models/terrace.zarr
clean_synthetic	Removes the lot

Every field is a closed-form expression over a 48 × 44 km patch of coast, so the data regenerates identically anywhere and a test can work out the expected answer by hand. nzcvm/synthetic.py documents the world it describes.

The pieces are importable, so a test can use a field directly rather than going through a file:

from nzcvm import synthetic

synthetic.elevation(172.0, -43.6)  # 550.0 m above sea level, inland
synthetic.vs30(172.6, -43.6)  # 200.0 m/s, offshore

---

Code architecture

Four subpackages, each with a narrow responsibility.

nzcvm.models

Geospatial Rust wrappers and mesh I/O. MeshModel wraps one tetrahedral mesh. ModelTree combines many meshes into a priority-ordered BVH tree and handles alpha-composited queries. Surface interpolates values from a 2-D triangular surface mesh (used for the DEM and the Vs30 map). mesh provides the tetrahedral and structured mesh dataclasses and their I/O.

nzcvm.layers

A Layer accepts a Grid (an xarray Dataset of 3D coordinates) and a ModelRange, and returns Qualities. Layers chain via constructor injection (next_layer), and register themselves against a config class through an __init_subclass__ hook on Layer.

Built-in layers: QueryLayer, ElyLayer, OffshoreBasinLayer, CoastlineLayer, ClampLayer, BackusAveragedLayer.

build_pipeline assembles the chain from a config list, and execute_model_pipeline applies it to every grid with one map_blocks per grid. That hoists the chunked dispatch out of the layers: each layer always receives a fully concrete chunk and can use plain NumPy.

nzcvm.config

Grid and layer configuration. Each layer comes with a companion LayerConfig dataclass and each grid with a companion GridConfig, dispatched on the type discriminator. VelocityModelConfig is the top-level object and validates layer ordering and dependencies.

nzcvm.grids

Builds 3D curvilinear meshes (sw4, regular or emod3d) as xarray DataTree nodes, chunked lazily with Dask and assembled from a GridConfig by the build_grids_from_config single-dispatch function.

Qualities (in nzcvm.qualities) is an xr.Dataset subclass holding the typed velocity, density and quality-factor arrays returned by every layer.

---

Extending with custom grids, layers, and formats

Grids, layers, and output formats are all extension points.

Functional layers (simple case)

The @functional_layer decorator turns a plain function into a registered layer, generating a matching LayerConfig from its keyword parameters.

from nzcvm.grids.grid import Grid
from nzcvm.layers.core import Layer
from nzcvm.layers.functional import functional_layer
from nzcvm.query import ModelRange


@functional_layer
def scale_vs(
    grid: Grid,
    model_range: ModelRange = ModelRange.ALL,
    *,
    next_layer: Layer,
    factor: float = 1.0,
):
    """Multiply Vs throughout the model by *factor*."""
    qualities = next_layer(grid, model_range)
    qualities["vs"] = qualities["vs"] * factor
    return qualities

Add it to a config:

[[layers]]
type = "scale_vs"
factor = 0.9

See examples/near_fault.py for a fuller example that uses a spatial distance mask to perturb Vs near a fault zone.

Class-based layers

For layers that need state, caching, or a non-trivial config, subclass Layer and pass a matching LayerConfig via the config_cls keyword. A layer's __init__ receives (config, geometry, next_layer). The geometry is the domain footprint, useful for pruning resources at construction time.

from dataclasses import dataclass

import numpy as np
from shapely import Geometry

from nzcvm.config.layers.core import LayerConfig
from nzcvm.grids.grid import Grid
from nzcvm.layers.core import Layer
from nzcvm.qualities import Qualities
from nzcvm.query import ModelRange


@dataclass
class DepthFloorConfig(LayerConfig):
    """Raise Vs to a floor that grows linearly with depth."""

    surface_floor: float = 500.0
    gradient: float = 0.05  # m/s of floor per metre of depth
    type: str = "depth_floor"


class DepthFloorLayer(Layer[DepthFloorConfig], config_cls=DepthFloorConfig):
    def __init__(
        self, config: DepthFloorConfig, geometry: Geometry, next_layer: Layer
    ) -> None:
        super().__init__(config, geometry, next_layer)

    def __call__(
        self, grid: Grid, model_range: ModelRange = ModelRange.ALL
    ) -> Qualities:
        qualities = self.next_layer(grid, model_range)
        floor = self.config.surface_floor + self.config.gradient * grid.depth
        qualities["vs"] = np.maximum(qualities["vs"], floor)
        return qualities

Use it in TOML:

[[layers]]
type = "depth_floor"
surface_floor = 500.0
gradient = 0.05

Importing the module registers the layer, so import it before decoding the config.

Custom grid types

A grid is an xarray Dataset built through GridSchema, which fixes the contract every layer relies on: x, y, z and depth on the logical (i, j, k) index (metres, projected CRS, z positive down), plus the attributes below.

GridSchema fixes the data variables, but a builder can attach extra coordinates after construction, and every stage keeps them: layers, map_blocks, and the Zarr and NetCDF writers. The borehole grid labels its columns this way:

grid = grid.assign_coords(site=("i", ["GULL", "TERR"]))

The name must avoid RESERVED_COORDINATES, since a coordinate shadows a variable or attribute of the same name. The csv and parquet writers turn any coordinate outside the contract into a label column.

Here is a transect: a line of vertical columns between two points, shaped (n, 1, nk).

import numpy as np
import shapely

from nzcvm.grids.grid import Grid, GridSchema


def transect_grid(
    start: tuple[float, float],
    end: tuple[float, float],
    n: int,
    bottom: float,
    dz: float,
) -> Grid:
    """A line of vertical columns of query points, sampled at *n* stations."""
    depth = np.arange(0.0, bottom, dz, dtype=np.float32)
    along = np.linspace(0.0, 1.0, n, dtype=np.float32)
    x = start[0] + along * (end[0] - start[0])
    y = start[1] + along * (end[1] - start[1])

    # Broadcast the (n,) line and the (nk,) depths into (n, 1, nk).
    x, _ = np.meshgrid(x, depth, indexing="ij")
    y, depth = np.meshgrid(y, depth, indexing="ij")
    x, y, depth = (array[:, np.newaxis, :] for array in (x, y, depth))

    return GridSchema.new(
        x=x,
        y=y,
        z=depth,
        depth=depth,
        name="transect",
        resolution=dz,
        geometry=shapely.LineString([start, end]),
        origin_lon=np.float32(174.7762),
        origin_lat=np.float32(-41.2865),
        azimuth=np.float32(0.0),
        grid_azimuth=np.float32(0.0),
        bottom_left_lon=np.float32(174.7762),
        bottom_left_lat=np.float32(-41.2865),
    )

Pass the result straight to a pipeline:

from pathlib import Path

from nzcvm.config.layers.query import QueryLayerConfig
from nzcvm.layers.pipeline import build_pipeline

grid = transect_grid(
    start=(1_749_150.0, 5_428_150.0),
    end=(1_759_150.0, 5_428_150.0),
    n=3,
    bottom=500.0,
    dz=100.0,
)
pipeline = build_pipeline(
    grid.geometry,
    [QueryLayerConfig(model_path=Path("models"), model_globs=["*.zarr"])],
)
qualities = pipeline(grid)
print(qualities.vs.values[0].ravel())  # Vs down the first column, in m/s

To drive a grid from a config file, register a builder against a GridConfig subclass. build_grids_from_config is a functools.singledispatch function returning a dict[str, Grid]. One entry per grid, since an SW4 domain produces one mesh per refinement level.

from dataclasses import dataclass
from typing import Literal

from nzcvm.config.grids import GridConfig
from nzcvm.grids.builder import build_grids_from_config
from nzcvm.grids.grid import Grid


@dataclass
class TransectConfig(GridConfig):
    start: tuple[float, float]
    end: tuple[float, float]
    n: int
    bottom: float
    dz: float
    type: Literal["transect"] = "transect"


@build_grids_from_config.register
def _(config: TransectConfig) -> dict[str, Grid]:
    return {
        "transect": transect_grid(
            config.start, config.end, config.n, config.bottom, config.dz
        )
    }

Which makes this config valid:

[grid]
type = "transect"
start = [1749150.0, 5428150.0]
end = [1759150.0, 5428150.0]
n = 3
bottom = 500.0
dz = 100.0

nzcvm.grids.borehole is the built-in version of this pattern: a registered GridConfig whose builder reads its own sites and DEM, over Dask-backed coordinates.

Custom output formats

Register a writer with @register_format. It receives the populated velocity model and the output path. Any extensions let nzcvm generate infer the format from the output path. Without them, select the format with --format.

from pathlib import Path

from nzcvm.formats import register_format
from nzcvm.velocity_model import VelocityModel


@register_format("vs_summary", extensions=(".vs.txt",))
def to_vs_summary(velocity_model: VelocityModel, path: Path) -> None:
    with path.open("w") as f:
        for name, (_grid, qualities) in velocity_model.pairwise.items():
            f.write(f"{name}: mean Vs {float(qualities['vs'].mean()):.1f} m/s\n")

Pass supports_quantisation=True if the writer accepts a quantise_arrays keyword for --quantise.

Plugins

To make custom layers, grids, or formats available to nzcvm generate without importing them yourself, declare entry points in the package that provides them. NZCVM imports each group as it loads the matching subsystem:

Entry-point group	Module contents
nzcvm.layer_configs	LayerConfig subclasses
nzcvm.layer_impls	Layer subclasses or @functional_layer functions
nzcvm.grid_configs	GridConfig subclasses
nzcvm.grid_impls	@build_grids_from_config.register builders
nzcvm.formats	@register_format writers

# pyproject.toml of the plugin package
[project.entry-points."nzcvm.layer_configs"]
depth_floor = "nzcvm_depth_floor.config"

[project.entry-points."nzcvm.layer_impls"]
depth_floor = "nzcvm_depth_floor.layer"

Configs and implementations load separately, so reading a config doesn't import a plugin's runtime dependencies. A @functional_layer module defines both at once, so list it only under nzcvm.layer_impls. Its configs then decode once something imports nzcvm.layers, which nzcvm generate always does. A plugin that fails to import logs a warning rather than breaking NZCVM. nzcvm plugins lists every available layer, grid, and format, which package provides it, and any plugin that failed to load.