Gentoo Packages

# Copyright 1999-2026 Gentoo Authors
# Distributed under the terms of the GNU General Public License v2

EAPI=8

PYTHON_COMPAT=( python3_{12..14} )
# Select ROCm 10 target flags; dependency atoms below constrain the stack.
# verified 2026-08-30
ROCM_VERSION=10.0
# Build as Release so upstream defines NDEBUG. cmake.eclass blanks the
# build type's own flag set, and cmake/Dependencies.cmake strips -DNDEBUG
# from CXXFLAGS unless the type is exactly Release, so a RelWithDebInfo
# build keeps every debug-only assert live: torch then aborts on checks
# upstream compiles out, such as storage().use_count() == 1 in the
# generated autograd code. verified 2026-09-16
CMAKE_BUILD_TYPE=Release
inherit python-single-r1 cmake cuda flag-o-matic prefix rocm

MYPN=pytorch
MYP=${MYPN}-${PV}

# Build against upstream's pinned composable-kernel revision.
CK_COMMIT=5a74dec07a894484b9489d0c0e00cd3b52652d18
CK_P=composable_kernel-${CK_COMMIT:0:8}

FLASH_PV=2.7.4
FLASH_PN=flash-attention
FLASH_P=${FLASH_PN}-${FLASH_PV}
FLASH_ATT_URI="https://github.com/Dao-AILab/${FLASH_PN}/archive/refs/tags/v${FLASH_PV}.tar.gz -> ${FLASH_P}.gh.tar.gz"

DESCRIPTION="A deep learning framework"
HOMEPAGE="https://pytorch.org/"
SRC_URI="
	https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz -> ${MYP}.tar.gz
	rocm? (
		https://github.com/ROCm/composable_kernel/archive/${CK_COMMIT}.tar.gz
		-> ${CK_P}.tar.gz
	)
	cuda? (
		flash? ( ${FLASH_ATT_URI} )
		memefficient? ( ${FLASH_ATT_URI} )
	)
"

S="${WORKDIR}"/${MYP}

LICENSE="BSD"
SLOT="0"
KEYWORDS="~amd64 ~arm64"
IUSE="cuda cusparselt distributed fbgemm flash gloo kineto memefficient
	mimalloc mkl mpi nccl nnpack +numpy onednn openblas openmp qnnpack
	rocm xnnpack"
RESTRICT="test"
REQUIRED_USE="
	${PYTHON_REQUIRED_USE}
	mpi? ( distributed )
	gloo? ( distributed )
	?? ( cuda rocm )
	rocm? (
		|| ( ${ROCM_REQUIRED_USE} )
		memefficient? ( flash )
	)
	cusparselt? ( || ( cuda rocm ) )
	flash? ( || ( cuda rocm ) )
	memefficient? ( || ( cuda rocm ) )
	nccl? ( rocm )
"

# The one-argument Quantize API works with FBGEMM 1.4 and 1.7, but 1.7's
# new rounding path is untested; retain the build-verified 1.4 cap.
# verified 2026-09-02
# aotriton-0.13b now matches upstream and supplies ROCm 10's rocm7.15 shim.
# verified 2026-09-02
# torch._C lives here; block older pytorch copies before merge.
RDEPEND="
	${PYTHON_DEPS}
	!!<sci-ml/pytorch-2.13.0
	dev-cpp/abseil-cpp:=
	dev-cpp/gflags:=
	>=dev-cpp/glog-0.5.0:=
	>=dev-libs/cpuinfo-2025.11.14
	dev-libs/libfmt:=
	dev-libs/protobuf:=
	dev-libs/sleef
	sci-ml/onnx
	virtual/lapack
	cuda? (
		dev-libs/cudnn
		>=sci-ml/cudnn-frontend-1.12.0:=
		>=dev-util/nvidia-cuda-toolkit-12.9:=[profiler]
		cusparselt? ( dev-libs/cusparselt )
	)
	fbgemm? ( >=sci-ml/FBGEMM-1.4 <sci-ml/FBGEMM-1.5 )
	gloo? ( >=sci-ml/gloo-2025.06.04[cuda?,rocm?] )
	kineto? ( ~sci-ml/kineto-0.4.0_p20260323 )
	mimalloc? ( dev-libs/mimalloc )
	mpi? ( virtual/mpi )
	nnpack? (
		sci-ml/NNPACK
		dev-libs/pthreadpool
	)
	numpy? ( $(python_gen_cond_dep '
		dev-python/numpy[${PYTHON_USEDEP}]
	') )
	onednn? ( sci-ml/oneDNN )
	qnnpack? (
		sci-ml/gemmlowp
		dev-libs/pthreadpool
	)
	rocm? (
		nccl? ( >=dev-libs/rccl-6.3:= <dev-libs/rccl-11:= )
		>=dev-util/hip-6.3:=       <dev-util/hip-11:=
		>=dev-util/roctracer-6.3:= <dev-util/roctracer-11:=
		>=sci-libs/hipBLAS-6.3:=   <sci-libs/hipBLAS-11:=[rocsolver(+)]
		>=sci-libs/hipBLASLt-6.3:= <sci-libs/hipBLASLt-11:=
		>=sci-libs/hipFFT-6.3:=    <sci-libs/hipFFT-11:=
		>=sci-libs/hipRAND-6.3:=   <sci-libs/hipRAND-11:=
		>=sci-libs/hipSOLVER-6.3:= <sci-libs/hipSOLVER-11:=
		>=sci-libs/hipSPARSE-6.3:= <sci-libs/hipSPARSE-11:=
		>=sci-libs/miopen-6.3:=    <sci-libs/miopen-11:=
		>=sci-libs/rocBLAS-6.3:=   <sci-libs/rocBLAS-11:=
		>=sci-libs/rocRAND-6.3:=   <sci-libs/rocRAND-11:=
		>=sci-libs/rocSOLVER-6.3:= <sci-libs/rocSOLVER-11:=
		memefficient? ( =sci-libs/aotriton-bin-0.13*:= )
		distributed? ( >=dev-util/rocm-smi-6.3:= <dev-util/rocm-smi-11:= )
		cusparselt? ( >=sci-libs/hipsparselt-6.3:= <sci-libs/hipsparselt-11:= )
	)
	distributed? (
		!rocm? ( sci-ml/tensorpipe[cuda?] )
		dev-cpp/cpp-httplib:=
	)
	xnnpack? (
		>=sci-ml/XNNPACK-2024.11
		dev-libs/pthreadpool
	)
	mkl? ( sci-libs/mkl )
	openblas? ( sci-libs/openblas )
"

# AsyncMM.cu instantiates a cooperative SM90 GEMM whose tile-scheduler call
# matches cutlass 4.6 and later; 4.4.2 rejects it as too few arguments. 4.6.1
# and 4.7.0 both compile it, so the cap stops at the first untested major
# rather than pinning: that call has already changed once across cutlass
# releases. verified 2026-09-16
DEPEND="
	${RDEPEND}
	dev-cpp/nlohmann_json
	dev-libs/flatbuffers
	dev-libs/FXdiv
	dev-libs/pocketfft
	dev-libs/psimd
	sci-ml/FP16
	$(python_gen_cond_dep '
		<dev-python/pybind11-3.0.5[${PYTHON_USEDEP}]
		dev-python/pyyaml[${PYTHON_USEDEP}]
		dev-python/typing-extensions[${PYTHON_USEDEP}]
	')
	cuda? ( >=dev-libs/cutlass-4.6.1[tools(+)] <dev-libs/cutlass-4.8[tools(+)] )
	onednn? ( sci-ml/ideep )
	rocm? (
		>=sci-libs/hipCUB-6.3:=    <sci-libs/hipCUB-11:=
		>=sci-libs/rocPRIM-6.3:=   <sci-libs/rocPRIM-11:=
		>=sci-libs/rocThrust-6.3:= <sci-libs/rocThrust-11:=
	)
	qnnpack? ( dev-libs/clog )
"

# Rebase against the post-sed source tree and keep shared patch names while
# their context remains exact. All patches apply without fuzz. verified 2026-09-02
PATCHES=(
	"${FILESDIR}"/${PN}-2.14.0-unbundle_fmt.patch.xz
	"${FILESDIR}"/${PN}-2.14.0-unbundle_kineto.patch.xz
	# Rebased for the Torch_SOURCE_DIR rename.
	"${FILESDIR}"/${PN}-2.14.0-unbundle_pocketfft.patch.xz
	"${FILESDIR}"/${PN}-2.5.1-cudnn_include_fix.patch.xz
	"${FILESDIR}"/${PN}-2.14.0-cpp-httplib.patch.xz
	# cpp-httplib 0.56 made Params an insertion-ordered map, which no longer
	# converts to the multimap reference the interface returns.
	# verified 2026-09-16
	"${FILESDIR}"/${PN}-2.11.0-cpp-httplib-params.patch.xz
	"${FILESDIR}"/${PN}-2.5.1-glog-0.6.0.patch.xz
	"${FILESDIR}"/${PN}-2.7.0-glog-0.7.1.patch.xz
	# Use glog's public initialization API; the internal symbol does not link.
	# verified 2026-09-02
	"${FILESDIR}"/${PN}-2.13.0-glog-exception-init.patch.xz
	# Applies after the libaotriton-path rewrite below.
	"${FILESDIR}"/${PN}-2.12.0-aotriton-fixes.patch.xz
	"${FILESDIR}"/${PN}-2.8.0-rocm-minus-flash.patch.xz
	"${FILESDIR}"/${PN}-2.14.0-rocm-distributed-link.patch.xz
	# hipFile is discovered but unused by PyTorch's CUDA-only CUFILE path; keep
	# it optional until ROCm storage support exists. verified 2026-09-03
	"${FILESDIR}"/${PN}-2.14.0-rocm-hipfile-optional.patch.xz
	"${FILESDIR}"/${PN}-2.9.1-torch_cpu.patch.xz
	# Rebased around the new adjacent perfetto section.
	"${FILESDIR}"/${PN}-2.14.0-gentoo.patch.xz
	"${FILESDIR}"/${PN}-2.14.0-mimalloc.patch.xz
	# Preserve 2.14's expanded activity list while guarding Kineto.
	"${FILESDIR}"/${PN}-2.14.0-removekineto-pr178960.patch.xz

	# Skip the submodule check for tarballs with system/prestaged deps.
	# verified 2026-07-18
	"${FILESDIR}"/${PN}-2.13.0-prebuildsteps-tarball-guard.patch.xz

	# Route install-time header wrapping through DESTDIR. verified 2026-07-18
	"${FILESDIR}"/${PN}-2.13.0-wrap-headers-destdir.patch.xz

	# Remove MKL cluster libs from the public link interface and force GNU OpenMP
	# so downstreams need neither Cluster Edition nor libiomp5. verified 2026-05-08
	"${FILESDIR}"/${PN}-2.12.0-mkl-public-scrub.patch.xz
)

src_prepare() {
	# Decompress patches into T because eapply cannot read xz payloads.
	local p b i
	mkdir -p "${T}"/patches || die
	for p in "${FILESDIR}"/*.patch.xz; do
		b=${p##*/}
		xz -dc "${p}" > "${T}/patches/${b%.xz}" || die
	done
	for i in "${!PATCHES[@]}"; do
		b=${PATCHES[i]##*/}
		PATCHES[i]="${T}/patches/${b%.xz}"
	done

	if use cuda && ( use flash || use memefficient ); then
		mv "${WORKDIR}"/${FLASH_P}/* third_party/${FLASH_PN}/ || die
	fi
	filter-lto #bug 862672

	sed -i \
		-e 's|::fmt-header-only||' \
		c10/CMakeLists.txt \
		cmake/Dependencies.cmake \
		torch/CMakeLists.txt \
		|| die

	# tensorpipe is in system, not a build target of caffe2
	sed -e '/target_compile_options_if_supported(tensorpipe/d' -i cmake/Dependencies.cmake || die

	sed -i \
		-e '/add_subdirectory.*third_party/d' \
		CMakeLists.txt \
		cmake/Dependencies.cmake \
		cmake/ProtoBuf.cmake \
		aten/src/ATen/CMakeLists.txt \
		|| die

	# System cutlass lacks an optional CuTeDSL example; do not fail mirroring.
	# Drop if cutlass starts installing it.
	# verified 2026-08-08
	sed -i -e 's/message(FATAL_ERROR/message(STATUS/' cmake/FileMirroring.cmake || die

	# The bundled fmt targets do not exist with system libfmt.
	# verified 2026-07-18
	sed -i \
		-e '/target_compile_definitions(fmt.*FMT_NO_UNIQUE_ADDRESS/d' \
		cmake/Dependencies.cmake \
		|| die
	sed -i \
		-e "/EXPORT/s|DESTINATION lib)|DESTINATION $(get_libdir))|" \
		c10/cuda/CMakeLists.txt \
		c10/CMakeLists.txt \
		c10/hip/CMakeLists.txt \
		|| die

	sed -i \
		-e "s|}/lib|}/\${CMAKE_INSTALL_LIBDIR}|g" \
		-e "/set(__AOTRITON_LIB/s|lib/|\${CMAKE_INSTALL_LIBDIR}/|g" \
		cmake/External/aotriton.cmake \
		|| die

	# Logging.h is incompatible with -Wextra-semi.
	sed -i 's/-Wextra-semi//' cmake/public/utils.cmake || die

	cmake_src_prepare
	pushd torch/csrc/jit/serialization > /dev/null || die
	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
	popd > /dev/null || die

	# Prefix hardcoded paths after applying patches.
	hprefixify \
		aten/CMakeLists.txt \
		caffe2/CMakeLists.txt \
		cmake/Metal.cmake \
		cmake/Modules/*.cmake \
		cmake/Modules_CUDA_fix/FindCUDNN.cmake \
		cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
		cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
		cmake/public/LoadHIP.cmake \
		cmake/public/cuda.cmake \
		cmake/Dependencies.cmake \
		torch/CMakeLists.txt \
		CMakeLists.txt

	if use rocm; then
		sed -e "s:/opt/rocm:/usr:" \
			-e "s:lib/cmake:$(get_libdir)/cmake:g" \
			-i cmake/public/LoadHIP.cmake || die

		# Every find_package(Torch) consumer reads the installed LoadHIP,
		# which otherwise falls back to an unslotted LLVM path when the
		# consumer has no HIP_CLANG_PATH; default it to HIP's clang.
		# verified 2026-10-01
		sed -e "s:\${ROCM_PATH}/lib/llvm/bin:$(hipconfig -l):" \
			-i cmake/public/LoadHIP.cmake || die
		grep -qF "set(_hip_clang_dir \"$(hipconfig -l)\")" \
			cmake/public/LoadHIP.cmake || die "LoadHIP clang default not rewritten"

		# Drop when composable-kernel becomes a system dependency.
		sed -e "s:third_party/composable_kernel:../composable_kernel-${CK_COMMIT}:g" \
			-i aten/src/ATen/CMakeLists.txt || die

		# Keep the remaining gfx101x ISA guard missing upstream. bug #959808;
		# verified 2026-09-02
		pushd "${WORKDIR}/composable_kernel-${CK_COMMIT}" > /dev/null || die
		eapply "${T}"/patches/composable-kernel-5a74dec0-expand-isa.patch
		popd > /dev/null || die

		# Work around LLVM issue https://github.com/llvm/llvm-project/issues/100802.
		sed -e 's/std::memcpy/memcpy/g' -i torch/headeronly/util/Half.h || die

		ebegin "HIPifying cuda sources"
		FBCODE_BUILD_TOOL="buck" ${EPYTHON} tools/amd_build/build_amd.py || die
		eend $?
	fi
}

src_configure() {
	if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
		ewarn "WARNING: caffe2 is being built for CUDA compute capability 7.5."
		ewarn "This default may not be optimal for your GPU."
		ewarn ""
		ewarn "To target your GPU, set TORCH_CUDA_ARCH_LIST through Portage's package.env"
		ewarn "mechanism and re-emerge caffe2."
		ewarn ""
		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
	fi

	local mycmakeargs=(
		-DBUILD_CUSTOM_PROTOBUF=OFF
		-DBUILD_TEST=OFF
		-DLIBSHM_INSTALL_LIB_SUBDIR="${EPREFIX}"/usr/$(get_libdir)
		-DPython_EXECUTABLE="${PYTHON}"
		-DTORCH_INSTALL_LIB_DIR="${EPREFIX}"/usr/$(get_libdir)
		-DUSE_CCACHE=OFF
		-DUSE_CUDA=$(usex cuda)
		-DUSE_DISTRIBUTED=$(usex distributed)
		-DUSE_FBGEMM=$(usex fbgemm)
		-DUSE_FLASH_ATTENTION=$(usex flash)
		-DUSE_GFLAGS=ON
		-DUSE_GLOG=ON
		-DUSE_GLOO=$(usex gloo)
		-DUSE_ITT=OFF
		-DUSE_KINETO=$(usex kineto)
		# KleidiAI and MAGMA are not packaged.
		-DUSE_KLEIDIAI=OFF
		-DUSE_MAGMA=OFF
		-DUSE_MEM_EFF_ATTENTION=$(usex memefficient)
		-DUSE_MIMALLOC=$(usex mimalloc)
		-DUSE_MKLDNN=$(usex onednn)
		-DUSE_MPI=$(usex mpi)
		-DUSE_NCCL=OFF
		-DUSE_NNPACK=$(usex nnpack)
		-DUSE_NUMA=OFF
		-DUSE_NUMPY=$(usex numpy)
		-DUSE_OPENMP=$(usex openmp)
		-DUSE_PYTORCH_QNNPACK=$(usex qnnpack)
		-DUSE_PYTORCH_METAL=OFF
		-DUSE_ROCM=$(usex rocm)
		-DUSE_SYSTEM_CPUINFO=ON
		-DUSE_SYSTEM_EIGEN_INSTALL=ON
		-DUSE_SYSTEM_FP16=ON
		-DUSE_SYSTEM_FXDIV=ON
		-DUSE_SYSTEM_GLOO=ON
		-DUSE_SYSTEM_NVTX=ON
		-DUSE_SYSTEM_ONNX=ON
		-DUSE_SYSTEM_PSIMD=ON
		-DUSE_SYSTEM_PTHREADPOOL=ON
		-DUSE_SYSTEM_PYBIND11=ON
		-DUSE_SYSTEM_SLEEF=ON
		-DUSE_SYSTEM_XNNPACK=$(usex xnnpack)
		-DUSE_TENSORPIPE=$(usex distributed $(usex !rocm))
		-DUSE_UCC=OFF
		-DUSE_VALGRIND=OFF
		-DUSE_XNNPACK=$(usex xnnpack)
		-DUSE_XPU=OFF
		-Wno-dev
	)

	if use mkl; then
		mycmakeargs+=(-DBLAS=MKL)
	elif use openblas; then
		mycmakeargs+=(-DBLAS=OpenBLAS)
	else
		mycmakeargs+=(-DBLAS=Generic -DBLAS_LIBRARIES=)
	fi

	if use cuda; then
		# Permit CUDA device probing in the sandbox (bugs #867706, #926116).
		cuda_add_sandbox
		addpredict "/dev/char/"

		mycmakeargs+=(
			-DUSE_CUDNN=ON
			-DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-7.5}"
			-DUSE_NCCL=OFF # CUDA NCCL is not packaged; nccl controls RCCL.
			-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
			-DUSE_CUSPARSELT=$(usex cusparselt)
		)

		[[ -v CUDACXX ]] && export PYTORCH_NVCC="${CUDACXX}"

		if use flash; then
			export FLASH_ATTENTION_FORCE_BUILD="TRUE"
			export FLASH_ATTN_CUDA_ARCHS="${CUDAARCHS:-${TORCH_CUDA_ARCH_LIST:-7.5}}"
		fi

	elif use rocm; then
		export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"

		# LoadHIP assumes an unslotted LLVM path; use Gentoo's HIP clang.
		# verified 2026-07-18
		export HIP_CLANG_PATH="$(hipconfig -l)"

		if use memefficient; then
			export AOTRITON_INSTALLED_PREFIX="${ESYSROOT}/usr"
		fi

		mycmakeargs+=(
			-DUSE_NCCL=$(usex nccl)
			-DUSE_SYSTEM_NCCL=ON
			-DCMAKE_REQUIRE_FIND_PACKAGE_HIP=ON
			-DCMAKE_DISABLE_FIND_PACKAGE_hipsparselt=$(usex !cusparselt) # disable automagic
			-DUSE_ROCM_CK_SDPA=OFF # requires flash + aiter, works only on gfx90a/gfx942/gfx950
		)

		append-cxxflags -Wno-deprecated-declarations -Wno-unused-result -Wno-unused-value
	fi

	if use onednn; then
		mycmakeargs+=(
			-DMKLDNN_FOUND=ON
			-DMKLDNN_LIBRARIES=dnnl
			-DMKLDNN_INCLUDE_DIR="${ESYSROOT}/usr/include/oneapi/dnnl"
		)
	fi

	cmake_src_configure
}

src_compile() {
	PYTORCH_BUILD_VERSION=${PV} \
	PYTORCH_BUILD_NUMBER=0 \
	cmake_src_compile
}

python_install() {
	python_domodule python/torch
	mkdir "${D}"$(python_get_sitedir)/torch/bin || die
	mkdir "${D}"$(python_get_sitedir)/torch/lib || die
	mkdir "${D}"$(python_get_sitedir)/torch/include || die
	ln -s ../../../../../include/torch \
		"${D}$(python_get_sitedir)"/torch/include/torch || die # bug 923269
	ln -s ../../../../../bin/torch_shm_manager \
		"${D}"/$(python_get_sitedir)/torch/bin/torch_shm_manager || die
	ln -s ../../../../../$(get_libdir)/libtorch_global_deps.so \
		"${D}"/$(python_get_sitedir)/torch/lib/libtorch_global_deps.so || die
}

src_install() {
	cmake_src_install

	# Drop the bundled /usr/lib aotriton copy; use the system libdir copy.
	# verified 2026-07-18
	if use rocm && use memefficient; then
		rm -rf "${ED}"/usr/lib/libaotriton_v2* "${ED}"/usr/lib/aotriton.images || die
	fi

	# CMake leaks wheel-layout files into /usr; keep only torch._C in
	# site-packages because pytorch supplies the pure-Python package.
	# verified 2026-07-22
	# Require the extension glob to match: silently deleting torch._C would pass
	# install yet break imports. Its destination has changed across releases.
	local _c_ext=( "${ED}"/usr/_C.cpython-*.so )
	[[ -f ${_c_ext[0]} ]] || die "torch._C not found at ${ED}/usr/ -- upstream moved its install destination"
	local _torchdir="${D}$(python_get_sitedir)/torch"
	mkdir -p "${_torchdir}" || die
	cp -p "${_c_ext[0]}" "${_torchdir}/" || die
	local d
	for d in "${ED}"/usr/*; do
		case ${d##*/} in
			bin|include|lib|lib64|share) ;;
			*) rm -rf "${d}" || die ;;
		esac
	done
	# On lib64 profiles remove leaked /usr/lib symlinks; never do this when
	# lib is the real payload directory.
	if [[ $(get_libdir) != lib && -d ${ED}/usr/lib ]]; then
		find "${ED}"/usr/lib -mindepth 1 -maxdepth 1 ! -name 'python*' \
			-exec rm -rf {} + || die
	fi

	# pytorch reuses this configuration.
	insinto "/var/lib/${PN}"
	doins "${BUILD_DIR}"/CMakeCache.txt

	rm -rf python || die
	mkdir -p python/torch || die
	cp torch/version.py python/torch/ || die
	python_install
}