---

name: qualcomm description: "Build, export, test, and debug ExecuTorch Qualcomm QNN backend workflows, including QNN SDK setup, compile specs, model enablement, Buck/CMake parity, and intermediate-output accuracy triage." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

qualcomm

Use this sub-skill when the user mentions Qualcomm, QNN, AI Engine Direct, HTP, Hexagon, QNN SDK, Qualcomm Android SoCs, QNN delegate tests, QNN accuracy divergence, or QNN Buck/CMake parity.

First Route

• Build or SDK setup: read Qualcomm workflows.
• Export/lowering/quantization: read Qualcomm workflows.
• Device/x86 tests: read Qualcomm workflows.
• Buck-vs-CMake CI failures: read Buck/CMake parity.
• Per-layer accuracy divergence or intermediate outputs: read Troubleshooting.

Hard Prerequisites

QNN workflows require SDK/toolchain state that a generic CPU environment cannot prove. Confirm QNN_SDK_ROOT, Android NDK path, target SoC model, device serial/host if running on device, and build directory before issuing run commands.

Bundled Planner

Generate a safe command plan without touching SDKs or devices:

python scripts/plan_qnn_command.py --soc SM8750 --build-dir build-android --artifact-dir /tmp/qnn-artifacts --device SERIAL
python scripts/plan_qnn_command.py --soc SM8750 --x86 --compile-only

The planner prints command templates and required environment variables; it does not build, install, push to a device, or download SDKs.

Cross-Links

• Use ../llm-workflows/SKILL.md for Llama/LLM model-asset planning before QNN backend execution.
• Use ../profiling-debugging/SKILL.md for generic ETDump/Inspector work; return here for QNN-specific intermediate-output/debugger workflows.
• Use ../setup-build/SKILL.md for generic CMake/Python environment failures.