IR Pretty Printing

Pretty-printing utilities for NN.IR.Graph.

We use these functions when:

• a compiler pass produces a malformed graph and we want a clear, compact dump,
• a verifier rejects a graph and we want to see the exact node/parent structure,
• we want a quick visualization of a graph in GraphViz.

This module is intentionally not a stable serialization format. The IR itself evolves (new ops, new payload conventions, new invariants), and we want the freedom to change the pretty output without treating it as part of the public API.

PyTorch analogy:

• pretty is like printing an FX graph as a node list,
• toDot is like emitting a visualization of the dataflow graph.

References:

• GraphViz DOT language: https://graphviz.org/doc/info/lang.html
• PyTorch FX (for the mental model of an op-tagged graph): https://pytorch.org/docs/stable/fx.html

Render a parent id list in a compact form.

parents in TorchLean's IR are data dependencies: an edge p -> n means "node n consumes the value produced by node p."

We keep this concise because it shows up in logs; for more context you typically want the whole graph listing (or the .dot output).

def NN.IR.Node.prettyLine (n : Node) : String

One-line rendering of a node (useful for logs).

Plain text rendering

def NN.IR.Graph.pretty (g : Graph) : String

Render the graph as a simple line-per-node listing.

This assumes the usual IR invariant that nodes are in id/topological order, but it does not enforce it; if you need validation, run Graph.checkWellFormed / Graph.checkShapes first.

GraphViz .dot rendering

def NN.IR.Graph.toDot (g : Graph) : String

Emit a GraphViz .dot representation.

Edges point from a node to its consumers (i.e. from parent to child), matching the “dataflow” picture most ML frameworks use.

Usage:

• IO.FS.writeFile "g.dot" (Graph.toDot g)
• dot -Tpng g.dot -o g.png