Versions#

V3.0.0#

Multi-node multi-GPU (MNMG) pipeline — training now requires torchrun and the new gpu: multi / gpu: single config field. Legacy non-torchrun entrypoints are no longer supported, while single-GPU runs remain supported with gpu: single (or NPROC=1).
New config schema — the YAML/JSON config format has changed substantially. The FullConfig schema now groups all settings under paths and models[].hyperparameters.gnn / models[].hyperparameters.xgb. Old flat configs are not compatible and must be migrated.
Triton server version — the serving container is now nvcr.io/nvidia/tritonserver:26.04-py3. Artifacts generated by earlier training containers are not compatible with this server version.
kind: GNN_XGBoost_NP — Node Prediction is now a first-class kind option in the config rather than a separate container entrypoint.
Mask file placement — train/val/test mask files are now placed in edges/ (EP) or nodes/ (NP). The previous location is no longer supported.
Edge filename parser — the src/rel/dst parsing now always takes the first token as src and the last token as dst. Multi-token node type names (e.g., credit_card) in edge filenames are no longer supported.

Automated HPO — llm_tune.sh (LLM-guided iterative tuning) and grid_search.py (Cartesian grid search) ship inside the container.
LLM Explainability — fraud_pipeline_explained ensemble chains GNN → XGBoost Shapley Value Sampling → OpenAI-compatible LLM for human-readable fraud explanations. Works with any OpenAI-compatible endpoint (NVIDIA NIM, OpenAI, Ollama, etc.).
Multi-GPU horizontal inference scaling — serve.sh accepts NUM_GPUS and INSTANCES_PER_GPU and patches all config.pbtxt files atomically.
Parquet and ORC support — node, edge, label, and attribute files can now be CSV, Parquet, or ORC. Formats can be mixed within a dataset.
XGBoost external memory modes — batched, extmem, and memmap options for datasets that exceed CPU RAM.
Focal loss — focal_gamma and focal_alpha GNN hyperparameters for severe class imbalance.
Production threshold calibration — prior_test parameter corrects the decision threshold for the production class ratio.
Zero-downtime model updates — Triton model versioning support documented; serve.sh supports rolling updates.
inf_num_neighbors — separate neighbor sampling cap for inference, enabling deployment on graphs with high-degree nodes.

Edge Prediction
Builds a GNN Model
Train an XGBoost model to predict fraud scores
Uses Wholegraph and cuGraph to encapsulate graph storage and complexity efficiently
Support to use XGBoost modeling optionally without GNN embeddings

Node Prediction
Builds a GNN Model
Train an XGBoost model to predict fraud scores
Uses Wholegraph and cuGraph to encapsulate graph storage and complexity efficiently
Support to use XGBoost modeling optionally without GNN embeddings