NVIDIA Attribution traces distributed AI workload failures

NVIDIA Attribution examines failures in the context of a distributed AI workload. Its Blueprint gathers the context and evidence; the Service correlates events and returns a diagnosis. The result identifies affected nodes and retains supporting symptoms and evidence. Several plausible causes can produce an ambiguous diagnosis. NVIDIA warns that its quality depends on complete, good-quality telemetry: the logs and measurements available to the system.

The quickstart offers a sample before connecting live infrastructure. It requires a suitable shell, uv and a running Docker installation. After preparing the repository, users build and start the container containing the service and web interface, then check service health before running attribution on the supplied sample configurations. The interface shows outcomes from the running rules version. Refresh an already-open page after the run finishes. Reusing the SQLite database appends another run for each sample job; a different database path separates the history.

The rule-authoring guide explains a separate responsibility: teaching the system to recognize evidence patterns. An extractor turns matching input into a structured event. Missing evidence instead calls for a connector change; an unreadable input format needs parsing support. Rule changes are committed and shipped with the Service, not uploaded through a runtime interface. Inline test cases check matching and emitted events when the catalog loads. They do not test the complete attribution pipeline, and the guide says an end-to-end test layer for new extractors is not currently provided.