We have a Ubuntu 22.04.03 server with eight A100 GPUs. The Nvidia driver is 550.144.03, the glibc version is 2.35, and the Cuda compiler is 12.4.131.
When the user tried to execute /usr/lib/nsight-systems/host-linux-x64/QdstrmImporter he go the following error: ./QdstrmImporter: /usr/lib/nsight-systems/host-linux-x64/libssh.so: version `LIBSSH_4_9_0’ not found (required by /usr/lib/nsight-systems/host-linux-x64/libSshClient.so)
I have been searching but I can’t find anything related to this error, and I can’t find any in reference to libssh version 4.9.0 for Ubuntu 22.04.
So when you do the run, the profile is not automatically being converted to an .nsys-rep file, and so the user is calling the qdstrmimporter manually and that manual call is failing?
If you have the 12.4 cuda toolkit (from the compiler version) you have a super old version of Nsys. My first recommendation would be to download the lates from Nsight Systems | NVIDIA Developer which is 2025.1 and try with that version.
Because of the time it takes to orchestrate all of the pieces in the CUDA Toolkit, the version of Nsys on the webpage is always newer and more full featured.
The NSight Systems was installed as part of the Lambda Stack from Lambda Labs, which kept it to the older version. I resolved the issue by installing the latest Nsight from the Ubuntu repository by first searching the versions that are currently available and install the latest one.