安装 flash-attention

fa2

# flash-attention
export CMAKE_CXX_STANDARD=17
export CMAKE_CXX_FLAGS="-D_GLIBCXX_USE_CXX11_ABI=1"
export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0"
git checkout v2.8.3
export FLASH_ATTENTION_FORCE_BUILD=TRUE
MAX_JOBS=4 python setup.py bdist_wheel

• fa3

export CMAKE_CXX_STANDARD=17
export CMAKE_CXX_FLAGS="-D_GLIBCXX_USE_CXX11_ABI=1"
#export TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.9 9.0"
export TORCH_CUDA_ARCH_LIST="8.0 8.6 8.9 9.0"
git checkout v2.8.3
#export FLASH_ATTENTION_FORCE_BUILD=TRUE
export FAHOPPER_FORCE_BUILD=TRUE
 source /opt/rh/gcc-toolset-13/enable
 MAX_JOBS=16 python setup.py bdist_wheel