Rank Layout¶

TorchLens uses Graphviz dot for local-topology graphs and can switch to an internal pure-Python rank layout for graphs whose long-range edges make dot expensive. No Node.js layout dependency is required.

Engine Selection¶

layout="auto" is the default. It estimates layout cost from the render graph before invoking Graphviz:

num_nodes + sum(rank_span for edges with rank_span > 12)

The rank layout is selected above 20,000 cost units. Local 5k-node chains stay on dot; hub-like graphs with a small number of long edges switch to rank.

The render graph is occurrence-aware for repeated argument uses. If one parent tensor feeds multiple argument slots of the same child, such as x + x or torch.cat([x, x]), TorchLens draws one arrow per slot. Commutative ops keep those parallel arrows unlabeled; non-commutative ops label each arrow with its argument slot.

Requirements¶

• Graphviz installed for final DOT/SVG/PDF rendering.

Usage¶

log = tl.trace(model, x)
log.draw(layout="auto", path="graph", file_format="svg")

Use layout="dot" to force Graphviz or layout="rank" to force the rank layout.

Use SVG for very large graphs. PDF renderers can produce empty or impractically large output at high node counts.

Troubleshooting¶