Skip to navigation
Reference

NVIDIA Attribution

View as Markdown

NVIDIA Attribution is enabled by default in the Helm release, image build, Tilt development stack, and local E2E harness. ADI calls the Attribution Adapter, which queries workload and node evidence and calls the NVIDIA Attribution provider. The provider Deployment and Service are owned by Helm alongside the adapter. Follow the operator installation guide for the complete stack.

Source and release prerequisites

The build source is dsx-ai-factory/attribution. Authorized source access remains necessary; public CI receives no private provider credentials. Its explicitly degraded E2E lane does not qualify the default Attribution-enabled release.

All default build paths use commit 7d6dff168bc282f471828fc8468e02de0df17480, matching the protected qualification workflow. The generated SDK is now under service/service-app/nvidia_attribution_client, with the Python namespace nvidia_attribution_client. Source preparation also includes the root LICENSE required by the upstream service Dockerfile, including for older sparse caches.

The adapter creates sessions and reads results through V2, and ingests existing log formats through the provider’s V1 compatibility endpoint, which shares those sessions. Prometheus vector and matrix results are flattened into the supported prometheus_metrics format; samples must retain __name__ and the configured hostname label. This compatibility update does not enable the currently disabled metric collection queries. Provider-boundary tests and local image checks do not constitute deployed GPU qualification.

The adapter build temporarily installs the client from this pinned checkout as a named BuildKit context. Public source/client publication and conversion to a verified pinned dependency in pyproject.toml and uv.lock remain release prerequisites. No speculative package coordinate is used.

Default image publication targets linux/amd64. Both Attribution builds select amd64, including local loads; both Deployments select amd64 nodes. Native amd64 end-to-end qualification remains required, and arm64 Attribution qualification is pending. Emulated builds do not satisfy the native gate.

Build and configure

With authorized upstream access, the ordinary build includes both provider images:

make build-publish IMAGE_PREFIX="$IMAGE_PREFIX" IMAGE_TAG="$IMAGE_TAG"

For a pre-existing checkout, set ATTRIBUTION_SOURCE_DIR to its absolute path and ATTRIBUTION_SOURCE_URL to its exact origin. The resolver verifies the immutable commit, origin, clean state, and required files before building any images. Optional SSH deploy keys require an explicit GitHub SSH source URL; never embed credentials in a URL. A failed source fetch is an enabled-provider failure; it never automatically switches to OFF.

The installation values configure image registry/tag, workload and node log selectors, and disable slow analysis until epilog ingestion is qualified. The chart derives same-release provider and adapter URLs. For an external provider, set attribution-service.enabled=false and attribution-adapter.fact.url to its HTTP(S) endpoint. For an external adapter, set attribution-adapter.enabled=false and anomaly-detection-isolator.attribution_service_url explicitly. An empty explicit endpoint is invalid when its consumer is enabled.

For standalone ADI, configure attribution_service_url before startup; enabled ADI rejects a missing or invalid HTTP(S) endpoint. The adapter similarly requires fact_url and its generated client. Network errors during a request remain structured Attribution errors so recovery can continue with visible degradation.

Check provider /v2/healthz and adapter /healthz separately, then send a representative request through the Adapter API. Healthy Pods alone are not attribution evidence.

Explicit degraded OFF mode

To build without provider source, select OFF explicitly:

ATTRIBUTION_ENABLED=false make build-publish \
  IMAGE_PREFIX="$IMAGE_PREFIX" IMAGE_TAG="$IMAGE_TAG"

Apply attribution-disabled-values.yaml after the installation values for the corresponding degraded release:

anomaly-detection-isolator:
  attribution_enabled: false
attribution-adapter:
  enabled: false
attribution-service:
  enabled: false

For local development or E2E, use E2E_ATTRIBUTION_MODE=unavailable; the default is available. OFF omits provider images and workloads and returns no attributions. It exercises recovery without provider diagnosis and does not qualify Attribution. Upgrading the Helm release with this override removes its provider and adapter workloads; no separately applied provider manifest remains.