GraphSpec Model Catalog

Import for the graph-authored example models in NN.GraphSpec.Models.*. Their structure is itself a typed Chain or DAG.Model, so the same architecture can be lowered to TorchLean and its graph shape reasoned about. Pure reference specifications (Transformer, ViT, Mamba, S4, UNet, VAE/VQ-VAE/GAN, classical baselines) live in NN.Spec.Models and are not imported here; runnable scripts live in NN.Examples.Models.

The set is a coverage ladder, not a catalog:

1. mlp: the smallest sequential Chain with a typed parameter ABI.
2. classifier: a caller-supplied feature chain followed by flattening and a linear classifier.
3. residualLinear: a minimal DAG.Model with a real skip connection.

The first two are sequential chains and the third is DAG-native. To turn a chain into a zero-initialized DAG.Model, import NN.GraphSpec.Chain.ToDAG.Model and use LowerToDAG.Chain.toDAGModelZeroInit.

See also:

• NN/GraphSpec/README.md for the overall layout and motivation.
• NN.GraphSpec.Chain.Lowering for the sequential DSL and lowering helpers.
• NN.GraphSpec.DAG for the DAG term language and its primitives.