MNN conversion failing backend test for RT DETR model

[jasseeeem]

平台(如果交叉编译请再附上交叉编译目标平台):

Platform(Include target platform as well if cross-compiling):

Tried converting in:

• x84_64 ubuntu20.04
• Apple M1 Pro macOS 14.7.1

MNN and converter was built following the documentation without any additional flags.

Model files:

• ONNX
• MNN

The above RT DETR model for object detection gives the correct output for bounding boxes, scores and labels when run on CPU, but the bounding boxes are wrong when run on CUDA GPU (scores and labels are correct). The model also fails backendTest.out with the following logs:

Test forward type: 0
Tolerance Rate: 0.050000
Open Model /Users/jaseem/Downloads/rtdetr_r50vd.mnn
hw.cpufamily: 458787763 , size = 4
The device support i8sdot:1, support fp16:1, support i8mm: 0
Input: 640,640,3,1
precision=0 in main, 280
modeNum=1 in main, 285
stopOp.c_str()=s  in main, 290
165: 4079.000000 != 1200.000000
/model/decoder/TopK_output_0 - 1 is error

Are RT DETR models unsupported in MNN?

[jxt1234]

1. what's the version of MNN ?
2. It seems that MNN's CPU is right but CUDA is wrong?
3. For TopK, if input's value is very closely. Inference engine may output different index. It's better to construct real data.

[jasseeeem]

1. 2.9.0, tried with the latest version too
2. Yes, the bounding box output is correct on CPU, but wrong on GPU
3. I tried with real images as well and saw the same issue

[jxt1234]

ok. We will check it.

[jasseeeem]

Thanks @jxt1234!