RV1106: All public toolkit versions produce zero output (only internal compiler works)

[arthors]

Summary

YOLOv5n model compiled with ALL public rknn-toolkit2 versions (1.6.0, 2.2.0, 2.3.2) produces zero output on RV1106 (Luckfox Pico Ultra W). The same architecture compiled with Rockchip's internal compiler (factory demo model) works correctly.

Reproduction

1. Train YOLOv5n (ultralytics) on custom desktop UI dataset — mAP50 0.932
2. Export ONNX (opset 12, simplify=True)
3. Compile with rknn-toolkit2:

   rknn.config(mean_values=[[0,0,0]], std_values=[[255,255,255]], target_platform='rv1106')
   rknn.load_onnx(model='best.onnx')
   rknn.build(do_quantization=True, dataset='calib.txt')

4. Deploy to Luckfox — rknn_init() returns 0 but rknn_run() output tensor is all zeros

Versions Tested

Toolkit	Result
1.6.0+81f21f4d	runtime error on 2.3.2 runtime; Softmax unsupported on 1.6.0 runtime
2.2.0	zero output
2.3.2	zero output + REGTASK bit width errors

Working Reference

Factory demo model (yolov5.rknn, compiled with 1.6.0 internal commit 585b3edcf, Dec 2023) works correctly — init 35ms, infer 88ms, 51K+ non-zero bytes in output.

Similar Reports

• Incorrect Face Detection Output When Using int16 Quantization with Yunet on RV1106 #375 (Oct 2025): Same zero-output on RV1106 with INT16 quantization
• Luckfox forum: Same issue reported since Nov 2025, unresolved
• Our GitHub issue: How did you compile working RKNN models for RV1106? (mAP 0.932 model ready, stuck at compilation) yiex/visiong#1

Request

Could Rockchip provide the internal compiler build (585b3edcf) or a fix for the public toolkit's INT8 quantization on RV1106? This blocks all custom model deployment on the Luckfox ecosystem.

[arthors]

Definitive proof: tested simplest model (4 nodes: Conv+ReLU), compiled with 1.6.0+81f21f4d. rknn_init=0, rknn_run=0, output ALL ZEROS. Same version number as working demo (585b3edcf), different binary. Bug is in compiler, not model.