1 geometry_backend

geometry_backend

PyTorch3D geometry backend preflight and provenance.

ARIA-NBV keeps geometry calls on PyTorch3D’s public API. This module owns only the user-facing backend request, provider capability check, and provenance record used by smoke tests and experiment artifacts.

1.1 Classes

Name Description
GeometryBackendProvenance Serializable geometry backend provenance for run artifacts.

1.1.1 GeometryBackendProvenance

geometry_backend.GeometryBackendProvenance(
    requested_backend,
    resolved_backend,
    torch_version,
    torch_device,
    platform_system,
    platform_machine,
    pytorch3d_version,
    pytorch3d_url,
    pytorch3d_commit,
    pytorch3d_cuda_available,
    pytorch3d_cuda_error,
    mojo_available,
    mojo_import_error,
    mojo_operations,
    dispatch_policy,
    counters,
)

Serializable geometry backend provenance for run artifacts.

1.1.1.1 Attributes

Name Description
requested_backend Explicit or environment-parsed backend policy at collection time; auto requests capability-based
resolved_backend Preflight-selected cpu, cuda or mojo provider route. This choice alone does not prove which kernels
torch_version Installed Torch version string captured with this provenance record.
torch_device Public geometry tensor device selected for the resolved route: cuda for CUDA, otherwise cpu,
platform_system Collecting host operating-system identity from platform.system(), used to admit the
platform_machine Collecting host architecture from platform.machine(), paired with platform_system for provider
pytorch3d_version Installed PyTorch3D distribution version, or None when package-version metadata is unavailable; not
pytorch3d_url Provider repository URL read from installed direct-URL metadata, or None when that metadata has no
pytorch3d_commit Provider commit identity from installed VCS metadata, or None when absent; source identity is
pytorch3d_cuda_available Whether the public PyTorch3D CUDA capability probe succeeded at collection time; no statement about
pytorch3d_cuda_error Capability-probe failure detail, or None after a successful CUDA probe; unavailable CUDA is not
mojo_available Provider-reported Mojo availability at collection time; false when its status does not declare
mojo_import_error Provider-reported Mojo import failure text, or None when no textual error is supplied.
mojo_operations Provider-declared Mojo operation names in reported order. Empty means no valid declaration was
dispatch_policy Provider-reported dispatch-policy description, or None when unavailable; ARIA records it without
counters Snapshot of provider-reported integer execution counters at collection time; an empty mapping means

1.1.1.2 Methods

Name Description
to_dict Return a plain JSON-serializable mapping.
1.1.1.2.1 to_dict
geometry_backend.GeometryBackendProvenance.to_dict()

Return a plain JSON-serializable mapping.

1.2 Functions

Name Description
collect_geometry_backend_provenance Collect backend provenance for run metadata and diagnostics.
expected_pytorch3d_identity Return the pinned provider for the current or explicitly recorded host.
main Print backend provenance as stable JSON.
requested_backend_from_env Parse PYTORCH3D_BACKEND using the provider’s supported values.
resolve_geometry_backend Resolve a requested backend without mutating PyTorch3D dispatch policy.
resolve_geometry_device Resolve tensors at the public PyTorch3D geometry boundary.

1.2.1 collect_geometry_backend_provenance

geometry_backend.collect_geometry_backend_provenance(requested=None)

Collect backend provenance for run metadata and diagnostics.

1.2.2 expected_pytorch3d_identity

geometry_backend.expected_pytorch3d_identity(system=None, machine=None)

Return the pinned provider for the current or explicitly recorded host.

Recorded host fields let evidence admission use the same platform pins as runtime dispatch without confusing the comparing host with the measured host.

1.2.3 main

geometry_backend.main()

Print backend provenance as stable JSON.

1.2.4 requested_backend_from_env

geometry_backend.requested_backend_from_env(value=None)

Parse PYTORCH3D_BACKEND using the provider’s supported values.

1.2.5 resolve_geometry_backend

geometry_backend.resolve_geometry_backend(requested=None)

Resolve a requested backend without mutating PyTorch3D dispatch policy.

1.2.6 resolve_geometry_device

geometry_backend.resolve_geometry_device(value)

Resolve tensors at the public PyTorch3D geometry boundary.

Mojo kernels in the Apple Silicon provider consume CPU tensors. The provider owns dispatch; this function is the sole ARIA-NBV translation from backend policy to the public PyTorch3D tensor contract.

Back to top