optimization-analysis-deployment

# AIMET optimization, analysis, and deployment Use this sub-skill when the user asks why accuracy dropped after quantization, how to use QuantAnalyzer or visualization outputs, how to choose mixed precision, how to use compression, or how to validate/export artifacts for target inference. ## Read/run first - Read [API overview](../../references/api-overview.md) for compression, ONNX analysis, SeqMSE, AdaRound, and related signatures. - Read [workflows](../../references/workflows.md) for the accuracy-debugging, compression, GenAILab, and deployment loops. - Read [backend compatibility](../../references/backend-compatibility.md) to separate CPU-valid analysis from CUDA/provider, GenAILab, cluster, or target-runtime requirements. - Read [troubleshooting](../../references/troubleshooting.md) for accuracy, visualization, compression, GenAILab, credential, cluster, and export-artifact symptoms. - Run [inspect_export.py](../../scripts/inspect_export.py) on an export directory before handing artifacts to target tooling. - Route Qualcomm AI Hub, QAIRT, QNN, HTP, DLC, and SDK-command tasks to [qualcomm-sdk-deployment](../qualcomm-sdk-deployment/SKILL.md). ## Core workflow 1. **Establish baselines.** Compare FP32, high-bit-width, and target-bit-width QuantSim behavior before choosing remedies. 2. **Separate weight and activation sensitivity.** Raise one side's bit-width while holding the other side at target precision. 3. **Apply the right PTQ remedy.** Use BatchNorm folding/Cross-Layer Equalization/per-channel quantization/AdaRound/SeqMSE for weight issues; use range-setting, higher activation precision, or mixed precision for activation issues. 4. **Analyze sensitive layers.** Use QuantAnalyzer, per-layer sensitivity, histograms, and layer-output comparisons to identify the quantizer or op causing the drop. 5. **Use mixed precision deliberately.** Lite/manual/automatic mixed precision should be driven by measured sensitivity and target latency/accuracy constraints. 6. **Compress when the target requires lower MACs or memory.** Select compression ratios, apply Weight SVD/Spatial SVD/Channel Pruning, evaluate, then fine-tune if required. 7. **Validate export artifacts.** Confirm the ONNX model and AIMET encodings match before QNN/QAIRT/AI Hub or other target handoff. ## Deployment boundaries AIMET produces simulation/export artifacts. It does not prove target-device correctness by itself. For Qualcomm AI Runtime, QNN, QAIRT, HTP, DLC, or AI Hub tasks, first validate the AIMET export pair and then use [qualcomm-sdk-deployment](../qualcomm-sdk-deployment/SKILL.md) for the local SDK or AI Hub workflow. ## Data and runtime boundaries - Compression and accuracy examples often assume ImageNet-style evaluators; replace them with a tiny evaluator while developing. - Visualization needs compatible Bokeh/HoloViews/HVPlot dependencies. - LLM/VLM GenAILab recipe benchmarks are now covered by [genai-lab](../genai-lab/SKILL.md), but they still require explicit model/dataset/credential/budget approval before execution. - On-target execution requires external SDKs/devices or AI Hub credentials; local export validation is not target proof. ## Expected answer shape When answering optimization/deployment tasks, include: - the baseline comparison being made; - which quantizers/layers are being isolated; - the selected AIMET remedy and why it matches weight/activation sensitivity; - the evaluator/calibration data requirements; - compression or mixed-precision constraints; - exported artifacts and validation steps; - explicit backend or target-runtime assumptions.