NVIDIA/structured-data-models

What it solves

It provides a GPU-native framework for working with structured data (tabular and relational), offering a unified interface to access foundation models that can perform predictions without needing extensive custom training for every new dataset.

How it works

The library implements a set of PyTorch-compatible tensor subclasses that handle various data types—such as numerical, categorical, datetime, and text—directly on the GPU. It integrates several pre-trained foundation models (like TabICLv2, KumoTabular, and KumoRelational) and provides composable, GPU-accelerated preprocessing and postprocessing tools to streamline the workflow from raw data to model inference.

Who it’s for

Data scientists and ML engineers who work with structured datasets and want to leverage foundation models for tabular or relational learning while maximizing GPU acceleration to reduce data movement and increase speed.

Highlights

• Foundation Model Library: Includes reference implementations for tabular models (TabICLv2, KumoTabular, TabFM) and relational models (KumoRelational).
• GPU-Native Tensors: Specialized tensor types for structured data that maintain PyTorch compatibility.
• Efficient Inference: Supports in-context learning and key/value caching for fast prediction passes.
• Optimized Pipeline: Recommended integration with cudf to keep dataframe operations on the GPU.