Symlink libcuda.so.1 to libcuda.so stubs?

[danpetry]

Comment:

In the system /usr/lib folder we have e.g.

lrwxrwxrwx.   1 root root       12 Mar 14 17:25 libcuda.so -> libcuda.so.1
lrwxrwxrwx.   1 root root       20 Mar 14 17:25 libcuda.so.1 -> libcuda.so.550.54.14
-rwxr-xr-x.   1 root root 24330976 Mar 14 17:25 libcuda.so.550.54.14

meaning that an application can access the real so through those three different points.
It seems unlikely that an application will use libcuda.so.550.54.14, but we've found one (llama.cpp) which looks for libcuda.so.1 (while building against cuda 12.3) so we had to create the symlink as part of that feedstock. Should it be here?

[jakirkham]

Can you please share the build?

[danpetry]

It's here

But I've just been informed of cuda-compat, is that what the purpose of that is?

[kmittman]

This allows using an older NVIDIA driver with a newer CUDA Toolkit, with several caveats.
More information about "Forward Compatibility" https://docs.nvidia.com/deploy/cuda-compatibility/#forward-compatibility-title

[danpetry]

Ah ok. This isn't the use case I've got, though. So should the libcuda.so.1 to libcuda.so symlink be moved to this feedstock? (opposite way around to the symlinks in the system installation, but seems to make sense for this purpose).

[isuruf]

Is libcuda.so.1 needed for linking or loading?

[danpetry]

linking (compilation fails without the symlink in the host environment during build)