I am able to run Triton-server 21.12 (server version 2.17) docker image with TensorRT backend on a GPU that has NVIDIA driver version 470 installed. However I noticed that Triton-server 21.12 requires NVIDIA driver version 495 or later on the release notes to support CUDA 11.5.
I am wondering if I am not getting any issues and able to successfully run the docker image with an older driver because of CUDA minor version compatibility. The logs do not have any warnings or information that it uses minor version compatibility though ( I came across some instances among the issues posted here, where the logs print this as an error or a warning about minor version compatibility/ driver version).
I want to know if I am able to use Driver 470 because of CUDA minor version compatibility or due to some other reason.
And I like to get some insights on if it is safe to rely on minor version compatibility even with future triton-server releases, where the required NVIDIA driver version given on release notes is higher, but the installed NVIDIA driver version is equal or greater than the minimum required driver, for the version of CUDA toolkit being used by Triton-server.