110 lines
4.4 KiB
CMake
110 lines
4.4 KiB
CMake
cmake_minimum_required(VERSION 3.14)
|
|
project(qwentts-ggml LANGUAGES C CXX)
|
|
|
|
set(CMAKE_CXX_STANDARD 17)
|
|
set(CMAKE_CXX_STANDARD_REQUIRED ON)
|
|
|
|
# version.h: embed git commit hash into all binaries.
|
|
# runs on every build, only rewrites if the hash changed.
|
|
set(VERSION_OUTPUT "${CMAKE_CURRENT_BINARY_DIR}/version.h")
|
|
add_custom_target(version ALL
|
|
COMMAND "${CMAKE_COMMAND}" "-DSRC_DIR=${CMAKE_CURRENT_SOURCE_DIR}" "-DOUTPUT=${VERSION_OUTPUT}"
|
|
-P "${CMAKE_CURRENT_SOURCE_DIR}/tools/version.cmake"
|
|
BYPRODUCTS "${VERSION_OUTPUT}"
|
|
COMMENT "Checking git version"
|
|
)
|
|
|
|
# pthread: required explicitly on older glibc (< 2.34) where libpthread
|
|
# is not merged into libc. Modern distros link it implicitly but aarch64
|
|
# and older x86_64 toolchains need the explicit dependency.
|
|
find_package(Threads REQUIRED)
|
|
|
|
# Suppress MSVC fopen/sprintf deprecation warnings
|
|
if(MSVC)
|
|
add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
|
|
endif()
|
|
|
|
# Put executables and backend .so in the same directory (build root).
|
|
# Without this, ggml defaults to bin/ for .so but executables stay in root,
|
|
# and ggml_backend_load_all() can't find the backends at runtime.
|
|
set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})
|
|
set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})
|
|
|
|
# Audio tokenizer tensor names can exceed default GGML_MAX_NAME of 64
|
|
add_compile_definitions(GGML_MAX_NAME=128)
|
|
|
|
# Harden: mark fread/fwrite/etc with warn_unused_result on all platforms
|
|
if(NOT MSVC)
|
|
add_compile_definitions(_FORTIFY_SOURCE=2)
|
|
endif()
|
|
|
|
# CUDA architectures: cover Turing to Blackwell for distributed binaries.
|
|
# Users can override with -DCMAKE_CUDA_ARCHITECTURES=native for local builds.
|
|
if(NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
|
|
find_package(CUDAToolkit QUIET)
|
|
if(CUDAToolkit_FOUND AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
|
|
set(CMAKE_CUDA_ARCHITECTURES "75-virtual;80-virtual;86-real;89-real;120a-real;121a-real")
|
|
else()
|
|
set(CMAKE_CUDA_ARCHITECTURES "75-virtual;80-virtual;86-real;89-real")
|
|
endif()
|
|
endif()
|
|
|
|
# ggml as subdirectory, inherits GGML_CUDA, GGML_METAL, etc. from cmake flags
|
|
add_subdirectory(ggml)
|
|
|
|
# Shared compile options and ggml linkage
|
|
macro(link_ggml_backends target)
|
|
target_include_directories(${target} PRIVATE
|
|
${CMAKE_SOURCE_DIR}/src
|
|
${CMAKE_SOURCE_DIR}
|
|
${CMAKE_BINARY_DIR}
|
|
)
|
|
target_include_directories(${target} SYSTEM PRIVATE
|
|
${CMAKE_SOURCE_DIR}/ggml/include
|
|
)
|
|
if(MSVC)
|
|
target_compile_options(${target} PRIVATE /W4 /wd4100 /wd4505)
|
|
else()
|
|
target_compile_options(${target} PRIVATE -Wall -Wextra -Wshadow -Wconversion
|
|
-Wno-unused-parameter -Wno-unused-function -Wno-sign-conversion)
|
|
endif()
|
|
target_link_libraries(${target} PRIVATE ggml Threads::Threads)
|
|
if(TARGET ggml-base)
|
|
target_link_libraries(${target} PRIVATE ggml-base)
|
|
endif()
|
|
foreach(backend cpu blas cuda metal vulkan)
|
|
if(TARGET ggml-${backend})
|
|
get_target_property(CURRENT_BACKEND_TYPE ggml-${backend} TYPE)
|
|
if (CURRENT_BACKEND_TYPE STREQUAL "MODULE_LIBRARY")
|
|
# DL mode: backend is loaded at runtime via dlopen,
|
|
# skip all link-time deps.
|
|
continue()
|
|
endif()
|
|
target_link_libraries(${target} PRIVATE ggml-${backend})
|
|
endif()
|
|
endforeach()
|
|
add_dependencies(${target} version)
|
|
endmacro()
|
|
|
|
# Core library shared between binaries. Holds the shared infrastructure
|
|
# (error/log routing, future common helpers) that any binary linking the
|
|
# pipeline needs. STATIC because we have actual sources now.
|
|
add_library(qwen-core STATIC src/qt-error.cpp)
|
|
target_include_directories(qwen-core PUBLIC src)
|
|
target_link_libraries(qwen-core PUBLIC ggml)
|
|
|
|
# quantize: GGUF requantizer (BF16 -> K-quants), shared policy with
|
|
# omnivoice.cpp / acestep.cpp.
|
|
add_executable(quantize tools/quantize.cpp)
|
|
link_ggml_backends(quantize)
|
|
|
|
# qwen-codec : standalone codec CLI (codes <-> WAV via 12Hz tokenizer)
|
|
add_executable(qwen-codec tools/qwen-codec.cpp src/pipeline-codec.cpp)
|
|
target_link_libraries(qwen-codec PRIVATE qwen-core)
|
|
link_ggml_backends(qwen-codec)
|
|
|
|
# qwen-tts : full TTS pipeline (Talker LM + 12Hz tokenizer decoder).
|
|
add_executable(qwen-tts tools/qwen-tts.cpp src/pipeline-tts.cpp src/pipeline-codec.cpp src/prompt-builder.cpp src/talker-forward.cpp src/code-predictor-forward.cpp)
|
|
target_link_libraries(qwen-tts PRIVATE qwen-core)
|
|
link_ggml_backends(qwen-tts)
|