failure of TensorRT 8.2 when running fp16 model on GPU Tesla T4

[yflv-yanxia]

Description

We recently converted a model similar to NAFNet into a TensorRT model, but we encountered two issues:

1. The fp16 model's inference results are incorrect while the inference results of the fp32 model are correct.
2. The fp32 model occupies a large amount of storage space, more than 80 times that of the original ONNX model.
   The first issue is more urgent, and we hope you can assist us in resolving it. Thank you!

Environment

TensorRT Version: 8.2.2.1

NVIDIA GPU:Tesla T4

NVIDIA Driver Version:450.36.06

CUDA Version:11.0

CUDNN Version:8.0.0

Relevant Files

onnx model
test image
correct result image
incorrect result image
inference code of onnx

Steps To Reproduce

Run inference code of onnx, you could get the correct result image.
Then convert the onnx model into tensorrt fp32/fp16 model, inference this two models get the results.

[zerollzeng]

8.2 is really old(and we don't fix bug on old release unless really necessary), could you please try latest version?

BTW TRT 10 will be released very soon.

[ttyio]

closing since no activity for more than 3 weeks, thanks all!