fix: build musa

[yeahdongcn]

This PR enables building for MUSA (S4000).

[yeahdongcn]

@thxCode Could you please run the workflow to check if the build succeeds? Thanks!

[yeahdongcn]

The CMake version has been updated in the latest commit to align with our base image, Ubuntu 20.04.

[yeahdongcn]

The project hierarchy differs slightly from Ollama. Ollama injects its ext_server as a sub-target of llama.cpp, allowing ext_server to inherit all the compilation settings from it.

[thxCode]

yes, the box is not in-tree structure, so we can observe the compilation detail. I think we almost achieved the goal. can you add this and have a try?

diff --git a/llama-box/CMakeLists.txt b/llama-box/CMakeLists.txt
index 80dcd90..172e673 100644
--- a/llama-box/CMakeLists.txt
+++ b/llama-box/CMakeLists.txt
@@ -77,6 +77,16 @@ if (GGML_CANN)
                 ${CANN_INSTALL_DIR}/runtime/lib64/stub)
     endif ()
 endif ()
+if (GGML_MUSA)
+    set(CMAKE_C_COMPILER clang)
+    set(CMAKE_C_EXTENSIONS OFF)
+    set(CMAKE_CXX_COMPILER clang++)
+    set(CMAKE_CXX_EXTENSIONS OFF)
+
+    set(GGML_CUDA ON)
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_MUSA)
+endif()
 add_executable(${TARGET} main.cpp param.hpp ratelimiter.hpp utils.hpp)
 add_dependencies(${TARGET} patch)
 target_link_libraries(${TARGET} PRIVATE version common llava ${CMAKE_THREAD_LIBS_INIT})

based on your work, I tried it on my laptop, is it green?

[yeahdongcn]

Yes, I also tried switching the default C/CXX compiler to Clang locally, and the compilation finished without errors. However, it seems odd that GGML_MUSA is used directly. Is this acceptable to you? 😂

[thxCode]

Yes, I also tried switching the default C/CXX compiler to Clang locally, and the compilation finished without errors. However, it seems odd that GGML_MUSA is used directly. Is this acceptable to you? 😂

LGTM at present.

[yeahdongcn]

Cool! Please see the latest commit, thanks.

[yeahdongcn]

The MUSA build is now successful.