---

name: profiling-debugging description: "Profile and debug ExecuTorch execution with ETRecord, ETDump, Inspector, visualization, delegate debug data, intermediate outputs, and runtime/export triage." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

profiling-debugging

Use this sub-skill when the user asks about ExecuTorch profiling, ETDump, ETRecord, Inspector, delegate debugging, model visualization, memory-planning inspection, intermediate outputs, or runtime/export accuracy/performance triage.

Route Here For

• Enabling and collecting ETDump at runtime.
• Generating ETRecord during export and using it to map runtime events to graph/source/module hierarchy.
• Loading artifacts with executorch.devtools.Inspector.
• Comparing delegate and CPU outputs, inspecting numerical divergence, and deciding when to route QNN-specific per-layer debugging to ../qualcomm/SKILL.md.
• Understanding memory planning and visualization artifacts.

Fast Path

1. Ask which artifacts the user already has: .pte, ETRecord, ETDump, debug buffer, eager reference outputs, backend logs.
2. If no ETRecord exists and source model/export is available, regenerate export with ETRecord enabled before collecting runtime data.
3. If only ETDump exists, perform raw timing inspection but explain that source/module attribution may be limited.
4. Use profiling workflows and API reference.
5. For artifact sanity checks, run:

python scripts/inspect_etdump_summary.py --etdump path/to/run.etdp --etrecord path/to/model.etrecord

The helper checks file presence and Inspector import availability; it does not parse private model data unless the installed package supports it.

Cross-Links

• Export and ETRecord generation details: ../export-runtime/SKILL.md.
• QNN-specific intermediate output debugger: ../qualcomm/SKILL.md.
• Backend choice and fallback behavior: ../backend-selection/SKILL.md.