ad-layer-visualizer

To visualize a transformer decoder layer from an AutoDeploy FX graph dump, you select a dump directory and layer number. The tool extracts the layer subgraph by following SSA dataflow from weight references, producing a hierarchical DOT or PNG diagram.

Yes, you can annotate transformer layer diagrams with real GPU kernel names and durations from an nsys trace. The tool extracts per-layer kernel sequences from the trace and maps them onto the diagram nodes to show execution time.

You need graphviz installed to render the layer diagrams into DOT or PNG files. You also need nsys if you want to extract GPU kernel timings from trace files and overlay them onto the graph nodes for performance debugging.

Graph debugging extracts a single layer subgraph from SSA-form dumps by following the SSA dataflow starting from layer-specific weight references and residual boundaries. This isolates the targeted decoder layer structure from the full model graph.

The best way to debug layer-level performance is to map nsys trace kernels onto a visualized layer diagram. This highlights what actually ran on the GPU and how long it took, allowing you to inspect structural and performance bottlenecks simultaneously.

Yes, the tool supports rendering hierarchical layer diagrams as PNG files. It generates a layer JSON schema from the extracted SSA subgraph, which the renderer then uses to produce the visual DOT or PNG output next to your JSON outputs.