skainet-lang-dag/sk.ainet.lang.dag

Package-level declarations

Types

Builder that records symbolic nodes/values. It is definition-only: no tensors are allocated.

annotation class DagDsl

abstract class DagModule

A reusable sub-graph module that can be instantiated within a DagBuilder.

data class GraphNodeDefinition(val id: String, val operation: Operation, val inputs: List<GraphValue<*>>, val outputs: List<GraphValue<*>>, val attributes: Map<String, Any?> = emptyMap())

Logical node definition captured by the DSL before lowering to sk.ainet.lang.graph.ComputeGraph.

Immutable program emitted by the DSL.

Collecting OpSink that accumulates OpTrace records and can convert them to a GraphProgram on demand.

data class GraphValue<out T : DType>(val nodeId: String, val outputIndex: Int, val spec: TensorSpec)

Symbolic value flowing through the DAG DSL. Every value is produced by a node output.

Records a symbolic initializer hint (used only as metadata on TensorSpec).

Lightweight, allocation-free builder that mimics the shape/initializer style of the data DSL but produces only TensorSpec metadata for the DAG DSL.

Entry point for the DAG DSL.

Defines a reusable graph component (module).

DSL helper for using modules.

Converts a list of OpTrace records (captured during a tracing forward pass) into a GraphProgram DAG suitable for optimization passes.