Add an official CMake target to build llama.cpp as a shared library.

[keiVision]

Proposal

Add an official CMake target to build llama.cpp as a shared (libllama.so) and/or static (libllama.a) library with all the necessary internal symbols exported for use as a "real" library.
Key Points:

• The library should include all public and required internal symbols, such as common_, LLAMA_BUILD_, etc.
• A minimal public C++ (or C) API header should be provided and documented for external consumption.
• Downstream projects should only need to link to this library, without managing a list of .cpp files.

Benefits

• Simplifies integration with other languages and build systems.
• Reduces maintenance burden for anyone building wrappers or bindings.
• Prevents build issues related to missing symbols or duplicated objects.
• Unifies ecosystem—all external bindings could standardize on the same library interface.

Prior Art / Examples

• Many popular C++ projects (e.g., OpenCV, TensorFlow, ONNX, libtorch) provide ready-to-link shared/static libraries for external consumers.
• Some llama.cpp community PRs and forks (e.g., #3280) have already explored this.

[ggerganov]

Isn't this already supported:

cmake --build build --target llama

[keiVision]

I did build libllama.so exactly as described:

cmake --build build --target llama

But when trying to use it as a shared library for my own C++ or Go project (via CGO), some symbols like common_token_to_piece and common_tokenize are missing, and I get "undefined reference" errors during linking.

I double-checked with:

nm externals/llama.cpp/build/bin/libllama.so | grep common_token_to_piece

and the symbol is not present.

Right now, my the only workaround is to manually add common.cpp and related sources to my CMakeLists.txt.

[ggerganov]

The libcommon is not part of the libllama. The libcommon is only used by the examples and the tools in this repository and it is not required by the libllama. The only API that is currently available from libllama is in include/llama.h.

[keiVision]

Thank you for the clarification!
Now it makes sense: libcommon is internal and not meant to be exported in libllama.so.
If I need functionality from common.cpp (for example, tokenization helpers), I will copy the necessary source files or reimplement what I need.
If possible, please consider documenting this explicitly in the README or CMake instructions (if it’s not already there - I couldn't find it). This will help new users avoid confusion!

Thank you!