accuracy drop of tensorrt fp32 compared with onnx

[Lilium]

Thank you for your open source!

---

I have trained r101-rtdetr and evaluated in Nvidia Orin in my dataset, results as below:

Model | onnx | trt-fp32 | trt-fp16 | trt-fp16-int8
mAP | 70.7 | 66.1 | 65.8 | 65.4

It is weired that trt-fp32 have accuracy drop (-4.6) compared with onnx, and I use Polygraphy to dig into bad layers, almost are MatMul and Pow, can be seen in log:
step3-op16-fp32.txt

Have you observed this situation?