# Copyright 2026 Gentoo Authors # Distributed under the terms of the GNU General Public License v2 EAPI=8 DISTUTILS_EXT=1 DISTUTILS_SINGLE_IMPL=1 DISTUTILS_USE_PEP517=setuptools PYTHON_COMPAT=( python3_{10..14} ) inherit cuda distutils-r1 multiprocessing pypi DESCRIPTION="Flash Attention: Fast and Memory-Efficient Exact Attention (Python component)" HOMEPAGE="https://github.com/Dao-AILab/flash-attention" LICENSE="BSD" SLOT="0" KEYWORDS="~amd64" IUSE="cuda rocm" REQUIRED_USE="${PYTHON_REQUIRED_USE}" # shellcheck disable=SC2016 RDEPEND="${PYTHON_DEPS} $(python_gen_cond_dep ' sci-ml/einops[${PYTHON_USEDEP}] ') >=sci-ml/pytorch-2.13.0[${PYTHON_SINGLE_USEDEP},cuda?,rocm?] cuda? ( >=sci-ml/pytorch-2.13.0[flash] )" DEPEND="dev-libs/cutlass ${RDEPEND}" # shellcheck disable=SC2016 BDEPEND="${PYTHON_DEPS} dev-build/ninja $(python_gen_cond_dep ' dev-python/psutil[${PYTHON_USEDEP}] test? ( sci-ml/triton-bin[${PYTHON_USEDEP}] ) ')" PATCHES=( "${FILESDIR}/${PN}-respect-flags.patch" ) pkg_setup() { if use cuda; then if [[ -z "${NVCC_PREPEND_FLAGS}" ]] || ! grep -q -- "--threads" <<< "${NVCC_PREPEND_FLAGS}"; then ewarn ewarn "If this hangs your system, try adding '--threads 1' to NVCC_PREPEND_FLAGS. Or try" ewarn "lowering the number of make jobs. Example:" ewarn ewarn " mkdir -p /etc/portage/env /etc/portage/package.env" ewarn " echo 'NVCC_PREPEND_FLAGS=\"--threads 1\"' > /etc/portage/env/flash-attn" ewarn " echo 'sci-ml/flash-attn flash-attn' >> /etc/portage/package.env/flash-attn" ewarn fi fi python-single-r1_pkg_setup } src_prepare() { cuda_src_prepare distutils-r1_src_prepare # PyTorch >= 2.14 headers require C++20 (torch/all.h #errors otherwise, and # c10/util/intrusive_ptr.h uses operator<=>/std::strong_ordering). Must run # after PATCHES, whose context lines still say c++17. sed -i 's/c++17/c++20/g' setup.py || die } src_compile() { export FLASH_ATTENTION_FORCE_BUILD=TRUE "MAX_JOBS=$(makeopts_jobs)" if use cuda; then cuda_add_sandbox # nvcc only supports a limited range of host compilers, and torch's # extension builder refuses to run when the default one is too new. local cuda_bindir cuda_bindir="$(cuda_gccdir)" || die "cuda_gccdir failed" export CUDAHOSTCXX="${cuda_bindir}/g++" export NVCC_CCBIN="${CUDAHOSTCXX}" export CC="${cuda_bindir}/gcc" CXX="${CUDAHOSTCXX}" export BUILD_TARGET=cuda fi use rocm && export BUILD_TARGET=rocm distutils-r1_src_compile } src_test() { # The tests run on the GPU. local dev for dev in /dev/nvidia* /dev/kfd /dev/dri; do [[ -e ${dev} ]] && addwrite "${dev}" done addpredict /dev/char addwrite /proc/self/task local EPYTEST_IGNORE=( # ROCm backends. tests/test_flash_attn_ck.py tests/test_flash_attn_triton_amd.py ) local EPYTEST_DESELECT=( # Uses JITFunction.cache_hook, which Triton 3.8 removed. tests/test_rotary.py::test_compilation_count ) # Many cases are sized for 80 GB GPUs; report out-of-memory as a skip. mkdir -p "${T}/pytest-plugins" || die cp "${FILESDIR}/${PN}-oom-skip.py" "${T}/pytest-plugins/flash_attn_oom_skip.py" || die local -x PYTHONPATH="${T}/pytest-plugins${PYTHONPATH:+:${PYTHONPATH}}" # Only tests/; without a path pytest also collects csrc/cutlass's scripts. local EPYTEST_FLAGS="-p flash_attn_oom_skip tests" # The installed package, not the source tree, has the built extension. mv flash_attn flash_attn.src || die distutils-r1_src_test mv flash_attn.src flash_attn || die } EPYTEST_PLUGINS=() distutils_enable_tests pytest