YOLOv11-N, quantized w8a16, compiled for the RB3 Gen 2 (Qualcomm QCS6490). The job died at the context-binary step with exactly one piece of feedback: exit code 14. No layer name, no reason, nothing to search for.
The same model in w8a8 went through fine — 5.03 ms on the NPU, against 134.7 ms for the default TFLite path. So the "safer precision" option produced no binary at all, and the tool couldn't tell us why.
The compile job was jp1193rnp, back in early November. I filed an issue (qualcomm/ai-hub-models#364) asking for one thing: whatever diagnostic exists underneath, surface it. A bare exit code isn't an error message, it's a locked door.
Ten days later the maintainer pasted the line the toolchain had been swallowing:
[4294967295] has incorrect Value 68, expected >= 73
That's the whole story in one line. The RB3 Gen 2's DSP is a Hexagon v68. Quantizing to w8a16 makes the compiler emit int16×int16 matmul kernels — and v68 doesn't have one. Those kernels need v73 or newer. Nothing was wrong with the model, the calibration, or the export. A dtype choice in a config file had quietly asked the chip for an instruction its generation doesn't ship.
The workarounds are simple once you know: stay on w8a8 for this silicon, or bring a custom AIMET config that keeps matmuls out of int16. The maintainer also forwarded the "surface the diagnostic" request to the QAIRT team, which is where the message gets swallowed — so the issue did its job even though the answer took ten days.
If you deploy to Hexagon v68 parts, treat w8a16 as unsupported for anything with matmuls and default to w8a8. And when a vendor tool hands you a bare exit code, the real error almost always exists one layer down. Filing for it is worth the ten minutes.
Top comments (0)