Structured Data Models

A GPU-native library of foundation models, tensor subclasses, and data processors for structured data.

• Models: Reference implementations of structured data foundation models, including tabular models (TabICLv2, KumoTabular, etc), and relational models (KumoRelational), built on a unified interface with room for future model families.
• Tensor semantics: PyTorch-compatible tensor types for numerical, categorical, datetime, text, and relational data.
• Data processing: Composable, extensible, and GPU-accelerated preprocessing and postprocessing for structured data workflows.

Installation

The structured-data-models package is available from Python 3.11 and PyTorch 2.7 onwards. Install from the main branch:

pip install git+https://github.com/NVIDIA/structured-data-models.git

Model Families

Tabular Foundation Models:

• TabICLv2 from Qu et al.: TabICLv2: A Better, Faster, Scalable, and Open Tabular Foundation Model (ICML '26)
• KumoTabular from Qu et al.: NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction ('26)
• TabFM from Kong et al.: Introducing TabFM: A Zero-shot Foundation Model for Tabular Data ('26)

Relational Foundation Models:

• KumoRelational from Hudovernik et al.: KumoRFM-2: Scaling Foundation Models for Relational Learning (CoRR '26)

Quick Tour

from sklearn.datasets import load_breast_cancer

import sdm

df = load_breast_cancer(as_frame=True).frame

# A lossless, fully tensorized representation of the raw data on GPU:
table = sdm.TableTensor.from_pandas(
    df=df,
    stypes=sdm.infer_stypes(df, overrides={"target": "categorical"}),
    device="cuda",
)

# Access to a variety of pre-trained structured data models:
model = sdm.models.TabICLv2(device="cuda")

# Default in-context learning forward pass:
model(
    x_context=table[:300].drop_columns("target"),
    y_context=table[:300, "target"],
    x_query=table[300:].drop_columns("target"),
    num_estimators=8,
)

# Fit + Predict forward pass via key/value caching for fast inference:
model.fit(
    x=table[:300].drop_columns("target"),
    y=table[:300, "target"],
    num_estimators=8,
)
model.predict(table[300:].drop_columns("target"))
model.clear()

Additional examples are available in examples/. Benchmarks for reproducing reported results live in benchmark/.

Notice and Disclaimer

This software automatically retrieves, accesses or interacts with external materials. Those retrieved materials are not distributed with this software and are governed solely by separate terms, conditions and licenses. You are solely responsible for finding, reviewing and complying with all applicable terms, conditions, and licenses, and for verifying the security, integrity and suitability of any retrieved materials for your specific use case. This software is provided "AS IS", without warranty of any kind. The author makes no representations or warranties regarding any retrieved materials, and assumes no liability for any losses, damages, liabilities or legal consequences from your use or inability to use this software or any retrieved materials. Use this software and the retrieved materials at your own risk.

License

The NVIDIA-authored source code is licensed under the Apache License 2.0. Third-party software and separately distributed model assets are documented in THIRD_PARTY_LICENSES.md:

• sdm/models/tabiclv2/ contains code derived from TabICLv2 under the BSD 3-Clause License; its terms are distributed in sdm/models/tabiclv2/LICENSE.
• sdm/models/tabfm/ contains code derived from TabFM under the Apache License 2.0; its terms are distributed in sdm/models/tabfm/LICENSE.
• sdm/models/kumo/relational/NOTICE documents KumoRelational's reuse of TabICLv2-derived components.
• third_party/pytorch/ contains the BSD 3-Clause License for material adapted from PyTorch in CONTRIBUTING.md; its terms are distributed in third_party/pytorch/LICENSE.
• third_party/contributor-covenant/ contains the MIT License for Contributor Covenant version 1.4; its terms are distributed in third_party/contributor-covenant/LICENSE.