Future support of Plugin Shared Libraries and PluginV3

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Description

Plugin shared libraries Developer Guide :: NVIDIA Deep Learning TensorRT Documentation do not support V3 Plugins as stated in this note in the developer’s guide

Blockquote
Note:IPluginRegistry’s loadLibrary() (C++, Python) functionality, demonstrated below, is not supported for plugin-shared libraries containing V3 plugin creators (IPluginCreatorV3One). As a workaround, define the entry point IPluginCreatorInterface* const* getCreators() in your library and then query this to enumerate each plugin creator and register it manually using IPluginRegistry’s registerCreator() (C++, Python).

Is this support expected in the future, or should I abandon the use of plugin shared libraries if I upgrade to V3?

EDIT: there is a workaround given, but it’s not clear to me if the workaround has all the benefits of the original hook. For example, should I handle the deregistering from the plugin registry as well? or will TRT handle that? The plugin must be registered at a global scope now, instead of at the local runtime scope.

Environment

TensorRT Version: 10.5

Relevant Files

N/A

Steps To Reproduce

N/A

2

It appears that the NOTE in the documentation is out of date/or misleading. There isn’t any workaround necessary: getPluginCreators() should be replaced by getCreators(). Here is an example from the TensorRT github TensorRT/plugin/common/vfcCommon.cpp at release/10.4 · NVIDIA/TensorRT · GitHub

3

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