GraphSpec to Sequential Models

This module converts a GraphSpec Chain to a TorchLean NN.Seq when every primitive supplies a Layer representation. The conversion packages deterministic parameter initialization and the parameter ABI expected by the sequential trainer.

This is a partial adapter, not the general GraphSpec execution path. Chain.toProgram interprets every supported GraphSpec primitive through the operation-polymorphic TorchLean.Program interface, including primitives with no Layer. By contrast, ToSequential.toSeq returns Except String and rejects a chain at the first primitive without a sequential-layer view.

The adapter threads a primitive occurrence index through the chain. Deterministic initializers use that index to derive stable per-layer seeds; the index is not part of GraphSpec's mathematical semantics.

Related modules:

• NN.GraphSpec.Core for the core DSL and its general Chain.toProgram translation.
• NN/GraphSpec/README.md for the relationship between GraphSpec and the runtime model API.

For g : Chain ps σ τ:

• ToSequential.toSeq g constructs a sequential model when every primitive has a Layer;
• Chain.toProgram g constructs the general operation-polymorphic program.

def NN.GraphSpec.ToSequential.Internal.toSeqFrom {ps : List Spec.Shape} {σ τ : Spec.Shape} (g : Chain ps σ τ) (i : ℕ) : Except String (Runtime.Autograd.Model.Layers.Seq σ τ × ℕ)

Lower a chain to a TorchLean Seq, threading a “layer occurrence index”.

The index is incremented for primitives with countsAsLayer = true.

def NN.GraphSpec.ToSequential.toSeq {ps : List Spec.Shape} {σ τ : Spec.Shape} (g : Chain ps σ τ) : Except String (Runtime.Autograd.Model.Layers.Seq σ τ)

Try to lower a GraphSpec chain into a Runtime.Autograd.Model.Layers.Seq.

Use this when you specifically want the Seq wrapper for training ergonomics. If all you need is an executable program, prefer Chain.toProgram: it is the more general path and does not require every primitive to have a Layer view.