1 TensorRT was linked against cuBLAS/cuBLAS LT 11.6.1 but loaded cuBLAS/cuBLAS LT 11.5.4
When using the TensorRT inference model, you will sometimes see the following logs generated by TensorRT itself:
TensorRT was linked against cuBLAS/cuBLAS LT 11.6.1 but loaded cuBLAS/cuBLAS LT 11.5.4
This is because the cuda version in our system environment is wrong. When we download TensorRT, from the downloaded file name, for example
TensorRT-8.2.0.6.Windows10.x86_64.cuda-11.4.cudnn8.2
I thought that as long as I downloaded the cuda 11.4 version, I didn’t care about the smaller versions after that, so I downloaded cuda 11.4.1 when I downloaded it, but the cuBLAS version of cuda 11.4.1 is only 11.5.4.8, so How to know which cuda version has the cuBLAS version?
Find the historical version page of CUDA from Nvidia’s official website: https://docs.nvidia.com/cuda/archive/ ,
Find all the links of cuda 11.4, and click on each one to see its Release Notes, find CUDA cuBLAS in it, followed by the version number, as shown in the figure below
Therefore, the version we should download and install is cuda 11.4.2 instead of cuda 11.4.1, so in order to avoid problems (although I don’t know what problems will occur if the corresponding cuda version is not applicable), it is better to proofread slowly version by version. Until a suitable cuda version is found.