Fix ggml-cuda using a driver symbol in NO_VMM mode

[milot-mirdita]

I have integrated the ProstT5 protein language into Foldseek. Thanks a lot for the great library! I am upstreaming a few fixes for issues I found in ggml during the integration. I hope that it's okay to push the changes here and that they get synced at some point to the main ggml repo.

This fix should be simple. cuGetErrorString uses the driver API and should not be available if GGML_CUDA_NO_VMM is set.

[JohannesGaessler]

I don't follow your logic. How are these two things related?

[milot-mirdita]

Would you prefer to rename GGML_CUDA_NO_VMM to GGML_CUDA_NO_DRIVER?

As far as I can tell GGML_CUDA_NO_VMM is used to disable the only subsystem that relies on the libcuda.so, everything else relies only on the cuda runtime.

[JohannesGaessler]

Can you explain the software and hardware setup where such an option would be beneficial?

[milot-mirdita]

cudart exists as a static binary. The cuda driver exists only as a .so/.dll. Without the driver dependency I can build a (nearly) static binary that only depends on libc, that can be executed on e.g. a CPU-only system that doesn't have cuda installed. If I have to link against libcuda.so, the binary will not run on this system.

If you run ldd on the foldseek binary (https://mmseqs.com/foldseek/foldseek-linux-gpu.tar.gz), you will not see any cuda dependencies, despite it working fully with cuda.

[slaren]

You can also avoid adding a hard dependency to CUDA by loading the backend dynamically. This wouldn't require disabling any features.