diff --git a/ggml/src/CMakeLists.txt b/ggml/src/CMakeLists.txt index b4105c1b4..440d4f506 100644 --- a/ggml/src/CMakeLists.txt +++ b/ggml/src/CMakeLists.txt @@ -256,6 +256,7 @@ ggml_add_backend(METAL) ggml_add_backend(RPC) ggml_add_backend(SYCL) ggml_add_backend(Vulkan) +ggml_add_backend(MUSA) foreach (target ggml-base ggml) target_include_directories(${target} PUBLIC $ $) diff --git a/ggml/src/ggml-musa/CMakeLists.txt b/ggml/src/ggml-musa/CMakeLists.txt new file mode 100644 index 000000000..ba78bf764 --- /dev/null +++ b/ggml/src/ggml-musa/CMakeLists.txt @@ -0,0 +1,107 @@ +if (NOT EXISTS $ENV{MUSA_PATH}) + set(MUSA_PATH /usr/local/musa) +else() + set(MUSA_PATH $ENV{MUSA_PATH}) +endif() + +set(CMAKE_C_COMPILER "${MUSA_PATH}/bin/clang") +set(CMAKE_C_EXTENSIONS OFF) +set(CMAKE_CXX_COMPILER "${MUSA_PATH}/bin/clang++") +set(CMAKE_CXX_EXTENSIONS OFF) + +list(APPEND CMAKE_MODULE_PATH "${MUSA_PATH}/cmake") + +find_package(MUSAToolkit) + +if (MUSAToolkit_FOUND) + message(STATUS "MUSA Toolkit found") + + file(GLOB GGML_HEADERS_MUSA "../ggml-cuda/*.cuh") + list(APPEND GGML_HEADERS_MUSA "../../include/ggml-cuda.h") + + file(GLOB GGML_SOURCES_MUSA "../ggml-cuda/*.cu") + file(GLOB SRCS "../ggml-cuda/template-instances/fattn-wmma*.cu") + list(APPEND GGML_SOURCES_MUSA ${SRCS}) + file(GLOB SRCS "../ggml-cuda/template-instances/mmq*.cu") + list(APPEND GGML_SOURCES_MUSA ${SRCS}) + + if (GGML_CUDA_FA_ALL_QUANTS) + file(GLOB SRCS "../ggml-cuda/template-instances/fattn-vec*.cu") + list(APPEND GGML_SOURCES_MUSA ${SRCS}) + add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS) + else() + file(GLOB SRCS "../ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu") + list(APPEND GGML_SOURCES_MUSA ${SRCS}) + file(GLOB SRCS "../ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu") + list(APPEND GGML_SOURCES_MUSA ${SRCS}) + file(GLOB SRCS "../ggml-cuda/template-instances/fattn-vec*f16-f16.cu") + list(APPEND GGML_SOURCES_MUSA ${SRCS}) + endif() + + set_source_files_properties(${GGML_SOURCES_MUSA} PROPERTIES LANGUAGE CXX) + foreach(SOURCE ${GGML_SOURCES_MUSA}) + set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22") + endforeach() + + add_library(ggml-musa + ${GGML_HEADERS_MUSA} + ${GGML_SOURCES_MUSA}) + + target_link_libraries(ggml-musa PRIVATE ggml-base) + target_include_directories(ggml-musa PRIVATE . ..) + + # TODO: do not use CUDA definitions for MUSA + target_compile_definitions(ggml PUBLIC GGML_USE_CUDA) + + add_compile_definitions(GGML_USE_MUSA) + add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X}) + add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y}) + add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER}) + add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE}) + + if (GGML_CUDA_GRAPHS) + add_compile_definitions(GGML_CUDA_USE_GRAPHS) + endif() + + if (GGML_CUDA_FORCE_DMMV) + add_compile_definitions(GGML_CUDA_FORCE_DMMV) + endif() + + if (GGML_CUDA_FORCE_MMQ) + add_compile_definitions(GGML_CUDA_FORCE_MMQ) + endif() + + if (GGML_CUDA_FORCE_CUBLAS) + add_compile_definitions(GGML_CUDA_FORCE_CUBLAS) + endif() + + if (GGML_CUDA_NO_VMM) + add_compile_definitions(GGML_CUDA_NO_VMM) + endif() + + if (DEFINED GGML_CUDA_DMMV_Y) + add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility + endif() + + if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16) + add_compile_definitions(GGML_CUDA_F16) + endif() + + if (GGML_CUDA_NO_PEER_COPY) + add_compile_definitions(GGML_CUDA_NO_PEER_COPY) + endif() + + if (GGML_STATIC) + target_link_libraries(ggml-musa PRIVATE MUSA::musart_static MUSA::mublas_static) + else() + target_link_libraries(ggml-musa PRIVATE MUSA::musart MUSA::mublas) + endif() + + if (GGML_CUDA_NO_VMM) + # No VMM requested, no need to link directly with the musa driver lib (libmusa.so) + else() + target_link_libraries(ggml-musa PRIVATE MUSA::musa_driver) + endif() +else() + message(FATAL_ERROR "MUSA Toolkit not found") +endif() \ No newline at end of file