Aipc Toolkit

AIPC Toolkit

Trigger Phrases

Always activate this skill when the user mentions:

Conversion

• "convert model to qnn" / "qnn conversion" / "convert to qnn"
• "convert model to dlc" / "snpe conversion" / "convert to dlc"
• "convert onnx to qnn" / "convert onnx to snpe"
• "generate context binary" / "context bin" / "qnn context binary"
• "run qnn inference" / "snpe inference" / "qairt inference"

Operator Patching

• "operator not supported" / "unsupported operator"
• "patch operator" / "operator patch"
• "converter failed" / "conversion failed"
• "dry run failed" / "unsupported ops found"

Diagnostics

• "check htp" / "htp ready" / "htp check"
• "aipc diagnose" / "environment check"
• "detect target soc" / "qairt devinfo" / "detect device info"
• "fastrpc memory map" / "err 1002" / "smmu" / "context binary too large"
• "split model" / "model split" / "split unet" / "context binary size"

Project Setup

• "create aipc project" / "init aipc project" / "setup aipc project"
• "aipc init" / "aipc setup"

When to Use

Use this skill for Qualcomm QAIRT/QNN/SNPE model bring-up:

• Export model to ONNX
• Inspect ONNX I/O
• Convert to QNN or SNPE/DLC
• Quantize model
• Generate context binaries
• Run inference and validation
• Create QAIRT/AIPC project

Required Guardrails

• Run skill scripts from their original skill path unless explicitly noted
• Do not swap out QAIRT toolchains ad-hoc
• QAIRT_SDK_ROOT must be set
• QAIRT_TMP_DIR — set when /tmp is full: qnn-onnx-converter copies external-data ONNX weights to a temp directory before processing. On servers where /tmp is a small tmpfs, this fails with Failed to copy external data to the disk at: /tmp/... permission or space issue. Set QAIRT_TMP_DIR to a path with at least 2× the model's weight file size free:

  export QAIRT_TMP_DIR=/path/with/enough/space   # e.g. /workspace/qairt_tmp
  

  Check before any converter call: df -h /tmp — if Use% is 100%, set QAIRT_TMP_DIR.
• LONG-RUNNING JOBS — CRASH RESILIENCE (MANDATORY): Any operation expected to take >2 minutes (conversion, context binary generation, inference) MUST be launched as a detached systemd service so it survives agent session crashes:

  systemd-run --user --unit=aipc-<task> --description="<desc>" bash /path/to/script.sh
  

  • Write all output to a persistent log file (not stdout): exec >> /path/to/job.log 2>&1
  • Use idempotent skip-if-done guards: if [ ! -f output.so ]; then ...; fi
  • Poll progress: systemctl --user status aipc-<task>.service --no-pager + tail -20 log
  • NEVER use nohup ... & or tmux — both are killed when the agent session ends.
• On Windows, do not rely on Python arch detection — use OS-native arch commands
• On ARM64 Windows, platform.machine() returns AMD64 under x86_64 emulation. Prefer the minimal QAIRT device probe script for SoC detection. Avoid CIM/WMI when possible. Model library target must match the Python process arch, not the CPU arch:
  • x86_64 emulated Python → windows-x86_64 model DLL
  • ARM64 native Python → windows-aarch64 model DLL
  • ARM64X/CHPE hybrid QNN runtime DLLs (arm64x-windows-msvc) bridge x86_64→native HTP
• Cross-platform shell commands:
  • Python scripts via subprocess.run() — no shell quoting issues
  • Inference execution policy (MANDATORY):
    • Run inference via scripts/aipc wrapper only.
    • MANDATORY: You MUST use the onnxwrapper.py provided within the aipc-toolkit skill (scripts/onnxwrapper.py).
    • PROHIBITED: NEVER use or fallback to the onnxwrapper.py bundled with the QAIRT SDK or qai_appbuilder package (e.g., from site-packages/qai_appbuilder/onnxwrapper.py). The skill version contains critical patches for AIPC workflows.
    • Before any final inference run, perform wrapper artifact preflight:
      • print which QNN artifact will be selected for {ONNX_FILE}
      • remove or quarantine stale matched artifacts (platform-dependent: <model>.onnx.dll.bin on Windows, <model>.onnx.so.bin on Linux) before deploying a new one
      • in context-binary mode, prefer ONNX-matching deployment filename: <model>.onnx.dll.bin on Windows, <model>.onnx.so.bin on Linux
      • the wrapper discovers context binaries by appending platform-specific suffixes to the ONNX path; see onnxwrapper.py:_find_qnn_model_file() for the full candidate list
    • If remote target execution is configured (for example RETMOE_DEVICE_INFO is set in project config), you MUST skip local host inference runs.
    • In this mode, acceptance and validation MUST be executed on the remote target only.
    • Local host inference is not allowed as an acceptance substitute.
    • Do NOT call snpe-net-run, qnn-net-run, or raw backend CLIs directly for final inference/validation.
    • Linux ARM runtime-libs pinning (MANDATORY for remote acceptance):
      • Do not rely on implicit wrapper auto-resolution when multiple target toolchain lib folders exist.
      • Explicitly set QAI_QNN_LIBS_DIR to the intended runtime directory for deployment and prepend it to LD_LIBRARY_PATH.
      • Example:

        export QAI_QNN_LIBS_DIR="$QAIRT_SDK_ROOT/lib/aarch64-oe-linux-gcc11.2"
        export LD_LIBRARY_PATH="$QAI_QNN_LIBS_DIR:$LD_LIBRARY_PATH"
        export ADSP_LIBRARY_PATH="$QAIRT_SDK_ROOT/lib/hexagon-v${DSP_ARCH}/unsigned"
        

      • If not pinned, mixed runtime loading may produce:
        • Failed to load skel, error: 4000
        • Transport layer setup failed: 14001
        • Failed to parse platform config: 14001
        • Unsupported SoC model ... / Invalid dsp arch ...
  • Avoid PowerShell inline pipelines and variables ($_, $env:, !) in command arguments passed via -Command (e.g., Where-Object { $_.LastWriteTime } gets expanded to { .LastWriteTime } by the outer shell, breaking syntax). Always write PowerShell pipelines to a temporary .ps1 file and execute it using the -File parameter, or use Python globbing/filesystem functions instead.
  • Console Encoding Guardrail: Local Windows shells can use non-UTF-8 encodings (such as cp950 or cp437), which frequently trigger UnicodeEncodeError or UnicodeDecodeError when processing console outputs with special characters. Always enforce UTF-8 encoding/decoding where possible, and use errors='replace' or errors='ignore' in Python subprocess handling.
• Escalation: If conversion still fails after export/patch/retry, do not silently replace model architecture. Record error + logs + ONNX snapshot → escalate with full bundle. For B3/B4/B7 criteria → open references/operator_patching.md.
• Dynamic-input ONNX: If ONNX has dynamic inputs, pass explicit shapes during conversion. See references/qnn_conversion.md (QNN: --input-dim) or references/snpe_conversion.md (SNPE: --source-model-input-shape).

⚠️ CRITICAL: Context Binary & Model Library Architecture (DO NOT SKIP)

Context binary (.dll.bin / .so.bin):

• ARM Windows: PREFERRED for fixed-SoC deployment — .dll (with ARM64X/CHPE runtime) also works.
• ARM Linux: OPTIONAL — .so works directly.
• The binary is SoC-specific and platform-independent — same .bin works whether host is x86 or ARM64.

Model library (.dll / .so):

• Must match the host process architecture, not the target CPU.
• On ARM64 Windows, the QAIRT Python venv runs under x86_64 emulation → qnn-model-lib-generator tool compile with -t windows-x86_64.
• Compiling for windows-aarch64 produces a DLL that the x86_64 emulated Python cannot load.

Context binary generation runs on the HOST (x86), not on the target device.

• The host uses qnn-context-binary-generator with soc_id/dsp_arch config to compile a binary for the target SoC.
• --model input must be a model library matching the host toolchain arch: windows-x86_64 on Windows, x86_64-linux-clang on Linux.
• The resulting .bin is then deployed to the target for inference.
• Do not add target cross-compilation steps unless the user explicitly asks for target model-library build.
• soc_id and dsp_arch are mandatory — identify them from the target device before generation.
• Do not use /sys/devices/soc0/soc_id — that is the Linux kernel ID, not the QAIRT soc_id.
• For how to read soc_id/dsp_arch from the target and full generation commands → open references/host_context_binary_gen.md.

For platform table, troubleshooting flow, and usage → open references/host_context_binary_gen.md.

ARM64X/CHPE note: On ARM64 Windows, QNN ships ARM64X hybrid DLLs (arm64x-windows-msvc/) that load from both x86_64-emulated and ARM64-native processes. The aipc wrapper + qai_appbuilder bundled libs use these automatically. Do not override QAI_QNN_LIBS_DIR or ADSP_LIBRARY_PATH to arm64x-windows-msvc unless the SDK version requires it.

Minimal device info probe: If CIM/WMI is unavailable, use registry + DriverStore inspection plus the QNN SoC enum table and qnn-platform-validator core-version output to resolve soc_id and dsp_arch. This is enough for context-binary generation on Snapdragon Windows hosts, and avoids device-specific hardcoding.

⚠️ CRITICAL: Operator Patching — Exhaustive Patching Required

Continue patching ALL unsupported ops until no replacement patterns exist. Never fall back to CPU. For patching rules, escalation policy (B3/B4/B7), and code templates → open references/operator_patching.md.

⚠️ CRITICAL: Project Setup Rules (DO NOT SKIP)

When the user requests "create project", "init project", or any project setup workflow for qairt: you must follow the AIPC skill end-to-end for all project setup actions. This is a required, certified workflow with defined acceptance criteria. Do not improvise, skip steps, or replace the workflow with manual setup. If this instruction is not followed exactly, the resulting work is considered invalid and must not be presented as compliant.

0. NEVER create local files, folders, or scripts yourself. Do not write any project files manually or spawn agents to do so.
1. ALWAYS run aipc_project_setup.py — never write AGENTS.md, CLAUDE.md, or aipc_plan.md manually.

   python /path/to/skills/aipc-toolkit/scripts/aipc_project_setup.py <project_dir>
   

2. Verify after the script: CLAUDE.md must be a symlink to AGENTS.md. Note: On Windows systems where symlink creation is restricted by local security policies (WinError 1314), the setup script's automatic copy fallback (copying AGENTS.md directly to CLAUDE.md) is fully acceptable and must NOT be treated as a setup failure.
3. Before auto-filling aipc_plan.md, inform the user that some Config values require their input (model name, target device, env script path, flow, etc.) and ask them to provide or confirm these before proceeding.
4. Then auto-fill derived and default values from the user's answers.
5. Never shortcut: manual file creation produces an incomplete scaffold (missing CLAUDE.md, wrong template, no sentinel). The script is the only correct path.

Quick Start

Bootstrap a project folder:

python skills/aipc-toolkit/scripts/aipc_project_setup.py path/to/project

This sets up:

• assets/aipc_AGENTS.md -> <project>/AGENTS.md
• <project>/CLAUDE.md linked to <project>/AGENTS.md
• assets/aipc_plan.md -> <project>/aipc_plan.md

Notes:

• If both AGENTS.md and CLAUDE.md already exist but are not linked together, setup stops with an error.
• If only CLAUDE.md exists, the script creates AGENTS.md as a symlink to CLAUDE.md before applying the AIPC agent content.

Then edit:

• aipc_plan.md Config section
• Placeholders in AGENTS.md / CLAUDE.md

Core Workflow

1. Setup QAIRT environment

   • Run aienv.ps1 (Windows) or source aienv.sh (Linux)
   • Verify QAIRT_SDK_ROOT is set

2. Export source model to ONNX

   • Use model's export script (e.g., export_onnx.py)
   • Recommended: opset_version=13 or higher

3. Inspect ONNX I/O and operator compatibility

   • Run: python aipc_inspect_onnxio.py model.onnx
   • Run converter dry-run to detect unsupported operators
   • If unsupported operators found → Proceed to Step 4

4. Operator Patching (if needed) — Approach Selection

   Decision Tree:

   1. Can you access PyTorch model BEFORE ONNX export?
      ├─ YES → Go to 2
      └─ NO  → Use ONNX Surgery (Approach 3)
   
   2. Is the operator an explicit PyTorch module?
      ├─ YES → In-Memory Module Replacement (Approach 1)
      └─ NO  → Go to 3
   
   3. Is the operator generated during ONNX export?
      ├─ YES → Custom Symbolic Handlers (Approach 2)
      └─ NO  → ONNX Surgery (Approach 3)
   

   Approach 1: In-Memory Model Patch (Preferred)

   • Modify model.forward() or replace module instances
   • Use references/operator_patching.md templates
   • Export patched model → model_patched.onnx

   Approach 2: Custom Symbolic Handlers (Excellent)

   • Register handlers before export: register_custom_op_symbolic()
   • Define ONNX graph for unsupported aten ops
   • Export with handlers active → model_patched.onnx

   Approach 3: ONNX Surgery (Fallback)

   • Use when source model is not accessible
   • Directly modify ONNX graph to replace unsupported ops
   • Validate: onnx.checker.check_model(model_patched.onnx)

   After Each Patch — Validation Gates:

   Gate	Check	Command	Pass Criteria
   1	ONNX Validity	onnx.checker.check_model()	No exceptions
   2	Converter	qnn-onnx-converter --dry_run	No unsupported ops
   3	Numerical	Compare with baseline	Cosine ≥ 0.95

   • Re-inspect: Run dry-run to verify no unsupported ops remain
   • Iterate: If new unsupported ops found → repeat Step 4
   • Track: Update aipc_plan.md with ALL patched operators
   • Stop when: All ops resolved OR exit criteria met (see Escalation Policy)

5. Convert float model

   • QNN path: python aipc_convert_fp.py --onnx model_patched.onnx ...
   • SNPE path: python aipc_convert_snpe.py --onnx model_patched.onnx ...

6. Optional: Quantization (INT8/INT16/A16W8)

   • Use aipc_convert_int.py (QAIRT default) or aipc_convert_aimet.py (AIMET path) with calibration data

7. Context binary generation (host-side)

• For ARM target context generation, do not run aipc_dev_gen_contextbin_x86.py without explicit target config.
  • Preferred: use direct qnn-context-binary-generator flow from references/host_context_binary_gen.md.
  • If using aipc_dev_gen_contextbin_x86.py, pass --config_file <backend_extension.json> (with soc_id / dsp_arch config).
• If generation fails (Windows): → Return to Step 4 (operator patching) — continue until no replacement patterns exist
• If generation fails (Linux): → Continue troubleshooting using references/host_context_binary_gen.md first.
  • You MUST attempt and record all applicable host-context methods before fallback:
    • confirm target soc_id/dsp_arch from device identity
    • run VTCM sweep (vtcm_mb=0,1,2,3,4,8)
    • test soc_id/dsp_arch alternatives from QAIRT mapping
    • if needed, test architecture-based config (htp_arch) / no-soc_id path
  • Only after all above methods fail with logs may you proceed with .so directly on Linux.
• See references/host_context_binary_gen.md for full commands and config templates

8. Inference + validation
   • Use aipc wrapper to run inference script
   • Validate accuracy against ONNX baseline.
   • Perform wrapper preflight before final run:
     • verify selected artifact path
     • clean stale matched .bin files in workdir (including .onnx.dll.bin on Windows, .onnx.so.bin on Linux)
     • ensure deployed context binary name matches ONNX discovery rule: <model>.onnx.dll.bin on Windows, <model>.onnx.so.bin on Linux
     • on Windows, the wrapper also matches <model>.onnx.dll.bin, <model>.dll.bin, <model>.dll in that order
   • On Linux ARM targets, set QAI_QNN_LIBS_DIR explicitly to intended runtime libs dir before running aipc.
   • If remote target execution is configured, you MUST run inference/validation on the remote target only.
   • In this case, skip local host inference entirely.
   • Final pass/fail must come from remote target results.

Reference Map

Open only what you need:

Script Index

⚠️ Inference must use scripts/aipc wrapper (including remote target runs). Direct snpe-net-run/qnn-net-run is for diagnostics only, not acceptance validation. ⚠️ Always prefer the wrapper scripts (aipc_convert_fp.py, aipc_convert_int.py, aipc_convert_aimet.py) over calling qnn-onnx-converter or qnn-model-lib-generator directly.