Is this page helpful?

Support Matrix#

Container#

Item

Value

Container image

nvcr.io/nvidia/cugraph/financial-fraud-training:3.0.0

Base image

NVIDIA cuGraph + PyG

CUDA version

12.6

Python version

3.10

Inference Server#

Item

Value

Triton Inference Server

nvcr.io/nvidia/tritonserver:26.04-py3

Triton HTTP port

8000

Triton gRPC port

8001

Triton metrics port

8002

Triton artifacts generated by v3.0.0 training are not compatible with earlier Triton server versions.

Supported GPU Architectures#

Architecture

Example GPUs

Supported

Ampere

A100 40GB, A100 80GB, A6000

Yes

Hopper

H100 80GB

Yes

Supported Data Formats#

Format

Extension

Training

Inference

CSV

.csv

Yes

Yes

Apache Parquet

.parquet

Yes

Yes

Apache ORC

.orc

Yes

Yes

NumPy (masks only)

.npy

Yes

—

Formats can be mixed within a single dataset. Mask files (train/val/test splits) are always .npy regardless of data format.

Supported Model Kinds#

kind

Description

GNN_XGBoost

Edge-level fraud prediction (default)

GNN_XGBoost_NP

Node-level fraud prediction

XGBoost

XGBoost-only baseline (no GNN); trains directly on node feature vectors

Supported GNN Encoders#

Encoder

encoder value

GraphSAGE

sage

Graph Attention Network

gat

Graph Transformer

transformer

GeneralConv

general

Supported LLM Providers (Explainability)#

Any OpenAI-compatible endpoint is supported:

Provider

Notes

NVIDIA NIM

Default; set NVIDIA_API_KEY

OpenAI

Set LLM_API_KEY, LLM_BASE_URL, LLM_MODEL

Azure OpenAI

Set LLM_API_KEY, LLM_BASE_URL, LLM_MODEL

vLLM (self-hosted)

Set LLM_API_KEY, LLM_BASE_URL, LLM_MODEL