How nvidia/structured-data-models works
Structured Data Models is a GPU-native Python library for structured-data tensors, preprocessing, and foundation models. The documented user workflow converts a dataframe into a typed TableTensor, supplies context and query data to a model for prediction, and can reuse a fit-time cache for later predictions. The graph also preserves the library’s relational-data, model-family, evaluation, explanation, and reusable neural-network capabilities. Some connections are documented only at a high level; unsampled implementations are represented by their existing directories rather than inferred internal call chains.
Main entry points: base.py Model interface · tabfm TabFM · table.py TableTensor · cache.py Inference cache
Components 22
Tensor semantics
- TableTensor
sdm/tensor/table.py - Typed tensors
sdm/tensor - Dataframe converters
sdm/tensor/io
Data processing
- Pre/postprocessors
sdm/processing - Processing recipes
sdm/processing/recipe.py
Model families
- Model interface
sdm/models/base.py - TabICLv2
sdm/models/tabiclv2 - TabFM
sdm/models/tabfm - KumoTabular
sdm/models/kumo/tabular - KumoRelational
sdm/models/kumo/relational - TimesFM3
sdm/models/timesfm3 - Classification evaluation
sdm/evaluation/classification.py - Model explanations
sdm/explain
Relational learning
- Relational data
sdm/relational/data.py - Relational tasks
sdm/relational/task.py - Relational joins
sdm/relational/join.py
Shared runtime
- Neural network layers
sdm/nn/attention.py - Inference cache
sdm/cache.py - Memory utilities
sdm/_memory.py - Model ensembles
sdm/ensemble.py
Other components
- Data scientist
- Input dataframe
Connections 24
- Data scientist → Input dataframe: provides data
README.md - Data scientist → TableTensor: constructs
README.md - Input dataframe → Dataframe converters: converts from
README.md - Dataframe converters → TableTensor: builds
sdm/tensor/table.py - TableTensor → Typed tensors: uses tensor types
sdm/tensor/table.py - Data scientist → Model interface: invokes prediction
README.md - Model interface → TabICLv2: unified interface
README.md - Model interface → TabFM: unified interface
README.md - Model interface → KumoTabular: unified interface
README.md - Model interface → KumoRelational: unified interface
README.md - Model interface → TimesFM3: model family
- Data scientist → Pre/postprocessors: preprocesses data
- Pre/postprocessors → Processing recipes: composes processors
sdm/processing/__init__.py - TableTensor → Relational data: combines with
- Relational tasks → Relational joins: uses joins
sdm/relational/task.py - Data scientist → Classification evaluation: evaluates predictions
- Data scientist → Model explanations: requests explanations
- Model interface → Inference cache: records fit state
README.md - TabFM → Neural network layers: uses layers
- TabFM → Memory utilities: checks memory
sdm/models/tabfm/cell_embedding.py - KumoTabular → Inference cache: uses cache
sdm/models/kumo/tabular/row_embedding.py - TabFM → Inference cache: uses cache
sdm/models/tabfm/row_embedding.py - TabICLv2 → Inference cache: uses cache
sdm/models/tabiclv2/row_embedding.py - Data scientist → Model ensembles: uses ensembles