Set LIB_PATH to libnvidia-ml.so.1 instead of libnvidia-ml.so on Linux

[dmitryduev]

In the official Go bindings for NVML, they use libnvidia-ml.so.1: https://github.com/NVIDIA/go-nvml/blob/0e815c71ca6e8184387d8b502b2ef2d2722165b9/pkg/nvml/lib.go#L30, and I believe the same is true for pynvml.

[scaronni]

Yes please, in driver 560 and above, as shipped in the CUDA repository, we've also removed the symlink to the unversioned library.

The approach of loading libnvidia-ml.so.1 is the correct approach. The unversioned library should be used only for linking building against it.

In the previous driver versions, there was in fact an nvidia-driver-devel subpackage which contained the unversioned libnvidia-ml.so library. But that was a mistake, as the package could not really be used as the unversioned libraries contained therein did not have any header for compiling against them and it was a leftover.

The few remaining unversioned libraries that are required in the driver have been moved to the main library packages.

Regarding NVML, the NVML stub and the headers are in the cuda-nvml-devel package, so if you need to link to it that's what should be installed.

Sample output for the RPM (deb is similar):

$ rpm -qpl cuda-nvml-devel-12-6-12.6.77-1.x86_64.rpm | grep targets
/usr/local/cuda-12.6/targets
/usr/local/cuda-12.6/targets/x86_64-linux
/usr/local/cuda-12.6/targets/x86_64-linux/include
/usr/local/cuda-12.6/targets/x86_64-linux/include/nvml.h
/usr/local/cuda-12.6/targets/x86_64-linux/lib
/usr/local/cuda-12.6/targets/x86_64-linux/lib/stubs
/usr/local/cuda-12.6/targets/x86_64-linux/lib/stubs/libnvidia-ml.a
/usr/local/cuda-12.6/targets/x86_64-linux/lib/stubs/libnvidia-ml.so

Again, please stick to loading libnvidia-ml.so.1 which is the correct approach. Thanks!

[dmitryduev]

@Cldfire can I please get a stamp?

[scaronni]

This is actually again #47