通过aten适配器引入flash attention库

[PanZezhong1725]

No description provided.

[PanZezhong1725]

已过期（适用commit d2aa36d）
复现方法：

1. 在third_party目录里拉取cutlass和flash_attn源码
2. 在third_party/flash-attention/csrc/flash_attn里添加如下CMakeLists.txt，注意替换其中路径

cmake_minimum_required(VERSION 3.18)
project(flash_attn LANGUAGES CXX CUDA)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_POSITION_INDEPENDENT_CODE ON)

# LibTorch include dirs
set(TORCH_INCLUDE_DIRS
    /home/panzezhong/Projects/InfiniCore/third_party/flash-attention/csrc/flash_attn/src
    /home/panzezhong/.conda/envs/myenv/lib/python3.13/site-packages/torch/include/torch/csrc/api/include
    /home/panzezhong/.conda/envs/myenv/lib/python3.13/site-packages/torch/include/
    /home/panzezhong/.conda/envs/myenv/include/python3.13/
    /home/panzezhong/Projects/InfiniCore/third_party/cutlass/include
)

# LibTorch libraries
set(TORCH_LIBS
    /home/panzezhong/.conda/envs/myenv/lib/python3.13/site-packages/torch/lib/libtorch.so
    /home/panzezhong/.conda/envs/myenv/lib/python3.13/site-packages/torch/lib/libtorch_cuda.so
    /home/panzezhong/.conda/envs/myenv/lib/python3.13/site-packages/torch/lib/libtorch_cpu.so
    /home/panzezhong/.conda/envs/myenv/lib/python3.13/site-packages/torch/lib/libc10.so
    /home/panzezhong/.conda/envs/myenv/lib/python3.13/site-packages/torch/lib/libc10_cuda.so
    /home/panzezhong/.conda/envs/myenv/lib/python3.13/site-packages/torch/lib/libtorch_python.so
)



# Collect all CUDA source files
file(GLOB FLASH_CU_SOURCES
    "${CMAKE_CURRENT_SOURCE_DIR}/flash_attn/src/*.cu"
)

add_library(flash_attn SHARED
    ${CMAKE_CURRENT_SOURCE_DIR}/flash_attn/flash_api.cpp
    ${FLASH_CU_SOURCES}
)
target_include_directories(flash_attn PRIVATE
    ${TORCH_INCLUDE_DIRS}
    ${CMAKE_CURRENT_SOURCE_DIR}/flash_attn
)

target_link_libraries(flash_attn PRIVATE
    ${TORCH_LIBS}
    /home/panzezhong/.conda/envs/myenv/lib/libpython3.13.so
    /home/panzezhong/.conda/envs/myenv/lib/libpython3.so
)

target_link_options(flash_attn PRIVATE "-Wl,--no-undefined")

set_target_properties(flash_attn PROPERTIES
    CUDA_ARCHITECTURES "80;86;90"
)

target_compile_options(flash_attn PRIVATE
    $<$<COMPILE_LANGUAGE:CUDA>:--expt-relaxed-constexpr>
    $<$<COMPILE_LANGUAGE:CUDA>:--use_fast_math>
)

add_definitions(-D_GLIBCXX_USE_CXX11_ABI=1)

3. 编译flash_attn

mkdir build
cd build 
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j8

编译成功后会在build目录里得到 libflash_attn.so，核心就是把它link到infinicore中去

4. 正常编译安装infinicore，运行mha_varlen算子测试

[PanZezhong1725]

更新自动化编译流程：

1. 设置cutlass路径环境变量 CUTLASS_ROOT
2. 配置环节打开 --aten 开关，并设置 --flash-attn 库位置
   xmake f --nv-gpu=y --ccl=y --cuda=$CUDA_HOME --aten=y --flash-attn=/home/panzezhong/Projects/InfiniCore/third_party/flash-attention -cv
3. flash attenion库会跟随infinicore_cpp_api一同编译安装