CUDA graph capture succeeds in Debug mode but fails in Release mode

[1250890838]

cudaError_t err = cudaStreamBeginCapture(context_->stream_, cudaStreamCaptureModeGlobal);
void** bindingsptr = bindingsPtr_.data();
bool execute_result = context->context_->enqueueV2(bindingsptr, context_->stream_, nullptr);
err = cudaStreamEndCapture(context_->stream_, &graph_);

when i run this code on release, After executing this statement
err = cudaStreamEndCapture(context_->stream_, &graph_)
the value of variable err is cudaErrorStreamCaptureInvalidated(901) and the value of variable execute_result is false.

but when i run this code on debug,everything is work.the value of variable err is cudaSuccess(0) and the value of variable execute_result is true.

Enviorment:

• TensorRT Version: 8.4.2.4

• NVIDIA GPU: RTX3060

• NVIDIA Driver Version: 472.12

• CUDA Version: 11.2

• Operating System: Windows10

how can i solve it? If you need more information, please let me know.

Thanks in advance!

[kevinch-nv]

Hard to say without seeing the rest of your code on what's going on.

Are you able to upgrade your CUDA / TRT versions? How are you linking to the TensorRT libraries in your application?

[1250890838]

i found the solution. At this link solution