DS 6 - Issue with re-registering a Yolo plugin

Hardware Details: Titan RTX dgpu
From NGC Docker container:
Deepstream 6
TensorRT 8.0.1-1
Cuda 11.3
Ubuntu 20.04
Nvidia Driver 470.86

I have a situation where I build a Deepstream pipeline that utilizes the Yolo plugin. My program creates an instance of the pipeline, and it runs just fine. However, if my program stops and destroys that pipeline, followed by recreating a new instance, upon setting the pipeline to “play” (i.e. running the pipeline), I get a Seg fault. The backtrace indicates that it occurred inside of libnvinfer.so, possibly while nvinfer is attempting to register the yolo plugin (again). If I run my app with a pipeline that doesn’t use yolo, I can destroy/recreate at will without issue. I’ve attached the backtrace
backtrace.txt (19.3 KB)
from gdb, and the config file used for this network.
yolo_face_config.txt (559 Bytes)

Could there be a problem with trying to re-register the same plugin a second time? If so, how can I work around this?


Sorry for the late response, is this still an issue to support? Thanks

Yes. Still an issue.

sorry for delay!

Yes, this a known issue, re-register the same plugin will cause TRT crash

Is there any way to “un-register” a plugin, or somehow reset TRT?

No for now.

You could add your plugin into TensorRT/plugin at main · NVIDIA/TensorRT · GitHub and build TRT plugin lib refer to deepstream_tao_apps/TRT-OSS/x86 at master · NVIDIA-AI-IOT/deepstream_tao_apps · GitHub

That’s disappointing.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.