SIGFPE Error Occurs When Converting TensorRT Model with Custom Plugins

Description

[11/01/2024-12:38:26] [V] [TRT] Formats and tactics selection completed in 38.437 seconds.
[11/01/2024-12:38:26] [V] [TRT] After reformat layers: 114 layers
[11/01/2024-12:38:26] [V] [TRT] Total number of blocks in pre-optimized block assignment: 119
[11/01/2024-12:38:26] [I] [TRT] Detected 2 inputs and 9 output network tensors.
[11/01/2024-12:38:26] [V] [TRT] [MemUsageChange] Subgraph create: CPU +18, GPU +0, now: CPU 2548, GPU 1538 (MiB)

Floating point exception (core dumped)

A Floating point exception (core dumped) occurs when I try to convert my onnx model to tensorRT.

It works fine with tensorRT 8.6.1, but not with tensorRT 10.4.0.

Environment

TensorRT Version: 10.4.0.0.26-1+cuda11.8
GPU Type: GeForce RTX 3090
Nvidia Driver Version: 560.28.03
CUDA Version: 11.8
CUDNN Version: 8.9.7
Operating System + Version: Ubuntu 20.04
Python Version: python 3.8
PyTorch Version: 1.10.2
Baremetal or Container: Baremetal

Relevant Files

plugins+model.zip (26.6 MB)

Steps To Reproduce

  1. Download plugins+model.zip.
  2. Extract all files and place it in the same directory.
  3. Copy and paste the following command.
trtexec --onnx=dsvt_full_simplified-trt.onnx \
--saveEngine=dsvt_full.trt \
--memPoolSize=workspace:4096 \
--verbose --plugins=libget_set.so --plugins=libwindow_partition.so \
--minShapes=pillar_features:1x128,voxel_coords:1x4 \
--optShapes=pillar_features:20000x128,voxel_coords:20000x4 \
--maxShapes=pillar_features:30420x128,voxel_coords:30420x4
  1. Wait a while and you can find the SIGFPE error.

Hello @jwher96 ,
Trting a repro at my end, and will share an update.

Thanks