ml_switcheroo.compiler.ir

Intermediate Representation (IR).

This module defines the language-agnostic graph data structures used to represent Deep Learning models after ingestion from source code (e.g. Python/LibCST) or explicit definition.

It acts as the contract between the Frontend (Ingestion) and the Backend (Synthesis).

Classes

LogicalNode	Represents a computation unit (Layer) in the graph.
LogicalEdge	Represents data flow between two nodes.
LogicalGraph	Language-agnostic representation of the neural network structure.

Functions

topological_sort(→ List[LogicalNode])

Sorts graph nodes by dependency order.

Module Contents

class ml_switcheroo.compiler.ir.LogicalNode

Represents a computation unit (Layer) in the graph.

id: str

Unique identifier (e.g. ‘conv1’).

kind: str

Operation type (e.g. ‘Conv2d’, ‘Input’, ‘Output’).

metadata: Dict[str, str]

Dictionary of configuration parameters (e.g. kernel_size=3).

class ml_switcheroo.compiler.ir.LogicalEdge

Represents data flow between two nodes.

source: str

Source node ID.

target: str

Target node ID.

class ml_switcheroo.compiler.ir.LogicalGraph

Language-agnostic representation of the neural network structure.

name: str = 'Model'

Name of the graph model/class.

nodes: List[LogicalNode] = []

Ordered list of nodes in the graph.

edges: List[LogicalEdge] = []

List of directed edges between nodes.

ml_switcheroo.compiler.ir.topological_sort(graph: LogicalGraph) → List[LogicalNode]

Sorts graph nodes by dependency order.

Ensures that for every edge u -> v, u appears before v in the returned list. Handles disconnected components and cycles gracefully by appending unreachable nodes in their original definition order.

Parameters:

graph – The logical graph to sort.

Returns:

List of nodes in execution order.