GraphSpec Model Catalog

Curated architecture import for GraphSpec users.

This file is the place a user should import when they are thinking “models as architectures.” It re-exports the pure model specifications from NN.Spec.Models and the graph-native examples in this directory.

We still keep the source files split by semantic layer:

• NN.Spec.Models.* contains pure mathematical/reference specifications such as Transformer, ViT, Mamba, S4, UNet, VAE/VQ-VAE/GAN, and classical baselines.
• NN.GraphSpec.Models.* contains graph-authored models whose structure is itself a typed Graph/DAG.Model, so we can compile the same architecture to TorchLean and reason about the graph shape.
• NN.GraphSpec.Models.TorchLean.* contains executable TorchLean constructors for models that are already useful as reusable autograd programs.
• NN.Examples.Models.* contains runnable scripts and training demos.

That split avoids circular dependencies and keeps proofs honest, while this umbrella gives users a single architecture-facing import that includes both the broad spec catalog and the graph-authored coverage ladder.

The current set is intentionally a coverage ladder, not an exhaustive catalog:

1. mlp: smallest sequential typed parameter ABI.
2. cnn2: sequential vision pipeline with convolution/pooling shape arithmetic.
3. residualLinear: minimal DAG model with a real skip connection.
4. ResNet18.model: larger DAG model with repeated residual blocks and projection shortcuts.

The examples intentionally mix two authoring styles, but they have one conceptual endpoint: DAG.Model.

• sequential Graph models for simple pipelines,
• DAG-native Model terms for residual / shared-structure examples.

NN.GraphSpec.Models is the single import for these GraphSpec-specific examples, regardless of which GraphSpec surface syntax they were authored in.

Included examples:

• NN.GraphSpec.Models.mlp (minimal sequential MLP) and NN.GraphSpec.Models.mlpDAGModelZeroInit (the same chain lowered to DAG),
• NN.GraphSpec.Models.cnn2 (sequential chain) and NN.GraphSpec.Models.cnn2DAGModelZeroInit (the same model, lowered to DAG),
• DAG-native models such as NN.GraphSpec.Models.residualLinear and NN.GraphSpec.Models.ResNet18.model.

See also:

• NN.GraphSpec/README.md for the overall layout and motivation.
• NN.GraphSpec.Core for the sequential DSL and lowering helpers.
• NN.GraphSpec.DAG for the canonical DAG IR and semantics.

If you are new to this directory, a good order is:

1. Models.mlp,
2. Models.cnn2,
3. Models.residualLinear as the minimal DAG/skip-connection example,
4. Models.ResNet18.model as the larger residual architecture.