Floating point exception of TensorRT 10.5 when converting an ONNX model with trtexec on GPU GeForce RTX 2080 Ti

[tom-brosch]

Description

I tried to convert an ONNX model of a U-Net to TensorRT using the trtexec command line tool.

Environment

TensorRT Version: 10.5

NVIDIA GPU: GeForce RTX 2080 Ti

NVIDIA Driver Version: 525.60.13

CUDA Version: 12.0

CUDNN Version: 9.6.0.74

Operating System: OpenSUSE

Baremetal or Container (if so, version): With TensorRT container (nvcr.io/nvidia/tensorrt:24.10-py3)

Relevant Files

Log file: logs.txt

Model link: model_cv1-4.zip

Steps To Reproduce

Commands or scripts:

• Start the docker container: nvidia-docker run -it --rm -v $PWD:/models nvcr.io/nvidia/tensorrt:24.10-py3
• Start the conversion: trtexec --onnx=/models/model_cv1-4.onnx --saveEngine=/models/model_cv1-4_trt.engine --minShapes=input:1x1x16x16x16 --optShapes=input:1x1x96x120x96 --maxShapes=input:1x1x96x120x96

Have you tried the latest release?: I also tried TensorRT 10.7 with the nvcr.io/nvidia/tensorrt:24.12-py3 image and it gives the same error.

Can this model run on other frameworks? I have the same issue when running the model through the TensorRT EP of ONNX Runtime.