Diagnose and fix MNN QNN backend errors, accuracy issues, and operator support on Qualcomm HTP.
npx skills add https://github.com/alibaba/MNN --skill qnn-debug
Please help me install this Agent Skill. Skill: qnn-debug Source: https://github.com/alibaba/MNN/tree/main/skills/qnn-debug Command: npx skills add https://github.com/alibaba/MNN --skill qnn-debug
Debugging the MNN QNN (Qualcomm HTP/NPU) backend is hard: runtime errors like 1002/6000/1003/6004, wrong inference results, fp16 precision gaps, and missing operator support all require deep backend knowledge. This Skill provides a systematic workflow to locate the first failing operator, distinguish real bugs from hardware fp16 accumulation, and add or adapt QNN operators.
Ask the assistant to debug why my MNN model gives wrong results on the QNN NPU backend and locate the first failing operator.
High-intent search queries and answers about installing and using this skill.
First establish a trusted CPU fp32 baseline with testMNNFromOnnx.py, then use MNN2QNNModel with --dump_intermediate_outputs to export all QNN intermediate tensors in one run. Compare each tensor against the CPU baseline to find the first operator whose input is good but output is bad.
Error 1002 means graphFinalize failed and 6000 means graphExecute failed. First try running with shapeMutable=false, which avoids a known input-copy issue. If it persists, enable QNN error logs in QNNBackend.cpp and check which operator violates constraints in HtpOpDefSupplement.html.
Compare QNN-fp16 against CPU-fp16, which acts as a well-behaved fp16 floor using fp32 accumulators. A sudden error jump at one operator where CPU-fp16 stays clean indicates a real bug, while smooth synchronized growth across both indicates fp16 accumulation.
First check SupportedOps.html, MasterOpDef.html, and HtpOpDefSupplement.html in the QNN SDK docs for input counts, parameter names, and HTP constraints. Then copy the closest existing operator template in source/backend/qnn/execution, register it in QNNUtils.cpp, and verify with the probe scripts.
The QNN online path does not copy model inputs in Session_Input_User mode (shapeMutable=true), so the first operator reads all zeros or graphExecute fails with error 6000. Setting shapeMutable=false routes input copying through Session_Input_Inside and fixes it without code changes.
The dump significantly increases graph outputs, memory usage, and execution time, so it is only for accuracy debugging, never for performance measurement or release builds. Existing finalized offline graphs cannot be retrofitted; you must regenerate a debug model with --dump_intermediate_outputs.
Pico-DeveloperConvert ONNX, TFLite, PB, and PyTorch models to QNN context binaries.
kngender5Quantize Colab models with AWQ, GPTQ, GGUF, and bitsandbytes.
limit5Deploy YOLO models to NPU hardware with quantization and accuracy validation.
alibabaOptimize MNN operators for ARM CPUs with multi-threading and assembly kernels.
