[Paddle-TRT] incorrect Plugin usage

[jeng1220]

The all data members of cloned plugin must be set in IPluginV2::clone. The plugin only calls IPlugin*::initialize once when creating ICudaEngine. The cloned plugin won't call IPlugin*::initialize after IExecutionContext is created, but TensorRT may clone the plugin during the creation of IExecutionContext and use the cloned plugin to infer the model.

I fix a plugin as example, see #24106. Following plugins have the same bug at least:

• ElementWisePlugin
• InstanceNormPlugin
• PReluPlugin
• SkipLayerNormPluginDynamic
• SplitPlugin

Read https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#ipluginext to know more.

[jeng1220]

@NHZlX
This is a serious bug. Please assign someone to fix it. Otherwise, all Paddle-TRT applications will crash after you upgrade TensorRT to next version, because new version of TensorRT always uses cloned plugin to infer the model

[NHZlX]

这个问题新版本lib拿到后会进行解决。

[paddle-bot-old]

Since you haven't replied for more than a year, we have closed this issue/pr.
If the problem is not solved or there is a follow-up one, please reopen it at any time and we will continue to follow up.
由于您超过一年未回复，我们将关闭这个issue/pr。
若问题未解决或有后续问题，请随时重新打开，我们会继续跟进。