GraphStormGNNInferenceEval

GraphStorm/DGL GNN out-of-time inference + evaluation on a custom GraphStorm GPU container (online-inference simulator → ROC-AUC / PR-AUC / Recall@Precision report + plots).

SageMaker step type	Processing
Node type	internal (consumes upstream, produces downstream)
Container entry point	graphstorm_gnn_inference_eval.py
Build-time requirement	none (SAIS SDK — fatal on load if absent)
Interface file	steps/interfaces/graphstorm_gnn_inference_eval.step.yaml

Compute

Compute kind	byo_container
Image	BYO container — config.image_uri is passed VERBATIM to AppSpecification.ImageUri / AlgorithmSpecification.TrainingImage (no image_uris.retrieve; the framework deps live in the image’s Dockerfile)
Container entrypoint	['bash', '/opt/ml/processing/input/code/entrypoint.sh'] (ContainerEntrypoint bypass — the image runs its own entrypoint instead of the SageMaker toolkit)

Functionality

GraphStorm GNN inference + out-of-time evaluation (ports Nexus run_evaluation.py): selects a checkpoint, converts multi-task→single-task if needed, auto-tunes GPU workers via live nvidia-smi VRAM probing, runs the online-inference simulator over the eval subgraph S3 folder, then computes ROC-AUC / PR-AUC / Recall@Precision and emits a report + plots. subgraph_s3_uri is a raw S3 folder streamed by the simulator (a job-arg, not a mounted channel).

Inputs (dependencies)

Input	Type	Required	Compatible producers
model_input	model_artifacts	yes	GraphStormGNNTraining
eval_seeds	processing_output	yes	GraphSubgraphExtraction, CradleDataLoading

Outputs

Output	Type
eval_results	processing_output

Consumers (downstream steps)

No cataloged step lists this step as a compatible source (it may be a terminal/sink step, or consumed via a generic source name).

Framework requirements

Package	Version
graphstorm	>=0.4
dgl	==1.1.3
torch	==2.1.0
polars	>=0.20.0
scikit-learn	>=1.3.0
matplotlib	>=3.7.0
numpy	>=1.21.0

---

← Back to the Step Catalog