cmake: fix cli build when LLAMA_BUILD_SERVER=OFF

[AsbjornOlling]

The Problem

The problem is that building with -DLLAMA_BUILD_SERVER=OFF does not work, since llama-cli now depends on server-context (since #17824).

To reproduce the issue on current latest master (commit 5642667):

cmake -B build -DLLAMA_BUILD_SERVER=OFF
cmake --build build

It will produce the following error:

[ 95%] Building CXX object tools/cli/CMakeFiles/llama-cli.dir/cli.cpp.o
In file included from /home/asbjorn/Development/llama-cpp-rs/llama-cpp-sys-2/llama.cpp/tools/cli/../server/server-task.h:12,
                 from /home/asbjorn/Development/llama-cpp-rs/llama-cpp-sys-2/llama.cpp/tools/cli/../server/server-context.h:2,
                 from /home/asbjorn/Development/llama-cpp-rs/llama-cpp-sys-2/llama.cpp/tools/cli/cli.cpp:6:
/home/asbjorn/Development/llama-cpp-rs/llama-cpp-sys-2/llama.cpp/tools/cli/../server/server-common.h:7:10: fatal error: mtmd.h: No such file or directory
    7 | #include "mtmd.h"
      |          ^~~~~~~~
compilation terminated.
make[2]: *** [tools/cli/CMakeFiles/llama-cli.dir/build.make:76: tools/cli/CMakeFiles/llama-cli.dir/cli.cpp.o] Error 1
make[1]: *** [CMakeFiles/Makefile2:4013: tools/cli/CMakeFiles/llama-cli.dir/all] Error 2
make: *** [Makefile:146: all] Error 2

This build error is what has prevented utilityai/llama-cpp-rs from bumping the llama.cpp version for the past month.

The Fix

I fixed this by moving the if (LLAMA_BUILD_SERVER) conditional from tools/CMakeLists.txt into the part of tools/server/CMakeLists.txt that builds the actual server executable. So now if LLAMA_BUILD_SERVER is OFF, it will still build the server shared libs (which llama-cli depend on), but not the server executable itself.

I fixed this by moving the cli target into the LLAMA_BUILD_SERVER conditional.
If LLAMA_BUILD_TOOLS=ON and LLAMA_BUILD_SERVER=OFF, it will not build cli or server, but it will build the rest.

I have run the ci/run.sh testing suite to confirm that everything works as before.

AI Disclosure

The code changes in this PR were all made and verified by human hands and eyes, but LLM assistance was used for identifying the problem and suggesting several solutions (one of which is this solution).

[ngxson]

You can skip building llama-cli when build_server is disabled. I prefer that simple fix. You won't use cli for binding anyway

[AsbjornOlling]

You can skip building llama-cli when build_server is disabled. I prefer that simple fix. You won't use cli for binding anyway

We don't need llama-cli for bindings, but don't we need to have LLAMA_BUILD_TOOLS enabled in order to have multimodal support in the bindings? It seems like all of the mtmd stuff lives in tools/mtmd/.

I have tested setting LLAMA_BUILD_TOOLS=OFF in utilityai/llama-cpp-rs, and it indeed does result in a linking error. It was enabled in llama-cpp-rs when multimodal support was added in utilityai/llama-cpp-rs#790

[ngxson]

Have you read my suggestion above?

[AsbjornOlling]

Have you read my suggestion above?

Ah okay. I read it, but misunderstood it. Sorry about that. 😅

I implemented your suggestion now.