Bug: docker GGML_CUDA=1 make [on llama-gen-docs] fails since arg refactor

[bartowski1182]

What happened?

Error given is

./llama-gen-docs: error while loading shared libraries: libcuda.so.1: cannot open shared object file: No such file or directory

Since this was only recently added in #9308 I'm guessing that's to blame

I've been able to get around it by running:

RUN ln -s /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1
RUN LD_LIBRARY_PATH=/usr/local/cuda/lib64/stubs/:$LD_LIBRARY_PATH GGML_CUDA=1 make -j64
RUN rm /usr/local/cuda/lib64/stubs/libcuda.so.1

but I guess my question is just why does it need this library at all and why is only this one failing?

Name and Version

b3707 ubuntu

What operating system are you seeing the problem on?

Linux

Relevant log output

./llama-gen-docs: error while loading shared libraries: libcuda.so.1: cannot open shared object file: No such file or directory

[slaren]

The current docker images use cmake and shouldn't be affected.

[bartowski1182]

I use make locally, is there a reason to prefer cmake?

[slaren]

Only the cmake build has all the options and the best defaults. For local testing make is ok, but cross-compiling and more advanced options cmake is necessary. The images were changed to use cmake to be able to link dynamically, and to set better defaults for the CUDA archs.