ml_switcheroo.compiler.differ

Topological Diff Engine.

Analyzes differences between two LogicalGraphs to produce a patch plan. It supports: 1. Identifying deleted nodes (nodes in source but not in target). 2. Identifying replacements (new fused nodes). 3. Mapping new nodes to anchors in the original graph for placement.

Classes

PatchAction

Base patch operation.

DeleteAction

Instruction to remove a node.

ReplaceAction

Instruction to replace an anchor node with new logic.

GraphDiffer

Calculates transformation steps to migrate from Source Graph to Target Graph.

Module Contents

class ml_switcheroo.compiler.differ.PatchAction

Base patch operation.

node_id: str
class ml_switcheroo.compiler.differ.DeleteAction

Bases: PatchAction

Instruction to remove a node.

class ml_switcheroo.compiler.differ.ReplaceAction

Bases: PatchAction

Instruction to replace an anchor node with new logic. For the Differ context, ‘new_node’ is the LogicalNode from the target graph.

new_node: Any
input_vars: List[str]
output_var: str
is_init: bool = False
class ml_switcheroo.compiler.differ.GraphDiffer

Calculates transformation steps to migrate from Source Graph to Target Graph.

Assumption: - Optimization is mostly fusion/deletion. - New nodes (fused) are distinct by ID or metadata. - We anchor rewrites to the last node of a fused chain (the sink).

diff(source: ml_switcheroo.compiler.ir.LogicalGraph, target: ml_switcheroo.compiler.ir.LogicalGraph) → List[PatchAction]

Compare graphs and return list of actions.

Algorithm: 1. Identify Removed IDs: source_ids - target_ids. 2. Identify Added/Changed Nodes. 3. Match Added Nodes to Removed Nodes (Find Anchor).

  • Heuristic: If FusedNode depends on input X, and RemovedNode N depends on input X, they might be related.

  • Current Heuristic: Use provenance or manual mapping supplied by Optimizer?

  • Simpler Heuristic: Fused nodes usually REPLACE a subgraph. The “output” variable usually stays consistent flow-wise, or we replace the Sink of the subgraph.

For this implementation, we rely on the Optimizer naming convention or graph structure. Since graph optimization logic isn’t fully inspectable here, we implement a Diff strategy based on ID presence.

Strategies: - If ID is missing in Target -> DELETE. - If ID is new/fused in Target -> REPLACE.

  • We need an Anchor in Source to attach the Replacement.

  • We attach the Replacement to the first available deleted node that matches topology?

  • Better: If Optimizer produced FusedNode fused_c1, and c1 is deleted, c1 is anchor?