NVIDIA TensorRT not compatible with CUDA 11.1

My environment

CUDA 11.1.0
Cudnn 8.0.4.30
TensorRT 7.2.0.14
RTX 3080
Tensorflow 2.3.1 built from source in the mentioned env.

The latest version of TensorRT 7.2 supports only CUDA 11.0. It does not work properly with CUDA 11.1.
Optimizing yolov4 with TRT yield this error :

2020-10-15 14:46:29.374511: W tensorflow/stream_executor/platform/default/dso_loader.cc:59]
Could not load dynamic library 'libnvinfer.so.7'; dlerror: /usr/local/cuda-11.1/lib64/libnvrtc.so.11.0: 
version `libnvrtc.so.11.0' not found (required by /usr/local/TensorRT-7.2.0.14/lib/libnvinfer.so.7); 
LD_LIBRARY_PATH: :/usr/local/cuda-11.1/lib64:/usr/local/cuda-11.1/include:/usr/local/TensorRT-7.2.0.14:/usr/local/TensorRT-7.2.0.14/include:/usr/local/TensorRT-7.2.0.14/lib
2020-10-15 14:46:29.374533: F tensorflow/compiler/tf2tensorrt/stub/nvinfer_stub.cc:49] getInferLibVersion symbol not found.

I already tried older CUDA version up to 11.0.3 but in all cases the driver doesn’t recognize the device.

We have already invested in some RTX 3080 GPUs for compute purposes and it’s underwhelming that we cannot use TRT with this GPU!

Is there any solution to make TRT work with CUDA 11.1 or any date for a compatible TRT version?

1 Like