From 72e4c0bc86eea366d063d15b3d5611e8f82ffd05 Mon Sep 17 00:00:00 2001
From: V3n3RiX <venerix@koprulu.sector>
Date: Mon, 24 Mar 2025 00:31:53 +0000
Subject: gentoo auto-resync : 24:03:2025 - 00:31:53

---
 sci-ml/caffe2/Manifest                |   6 +-
 sci-ml/caffe2/caffe2-2.4.1-r5.ebuild  | 304 -------------------------------
 sci-ml/caffe2/caffe2-2.4.1-r6.ebuild  | 305 +++++++++++++++++++++++++++++++
 sci-ml/caffe2/caffe2-2.5.1-r10.ebuild | 331 ++++++++++++++++++++++++++++++++++
 sci-ml/caffe2/caffe2-2.5.1-r9.ebuild  | 330 ---------------------------------
 sci-ml/caffe2/caffe2-2.6.0-r1.ebuild  | 330 ---------------------------------
 sci-ml/caffe2/caffe2-2.6.0-r2.ebuild  | 331 ++++++++++++++++++++++++++++++++++
 7 files changed, 970 insertions(+), 967 deletions(-)
 delete mode 100644 sci-ml/caffe2/caffe2-2.4.1-r5.ebuild
 create mode 100644 sci-ml/caffe2/caffe2-2.4.1-r6.ebuild
 create mode 100644 sci-ml/caffe2/caffe2-2.5.1-r10.ebuild
 delete mode 100644 sci-ml/caffe2/caffe2-2.5.1-r9.ebuild
 delete mode 100644 sci-ml/caffe2/caffe2-2.6.0-r1.ebuild
 create mode 100644 sci-ml/caffe2/caffe2-2.6.0-r2.ebuild

(limited to 'sci-ml/caffe2')

diff --git a/sci-ml/caffe2/Manifest b/sci-ml/caffe2/Manifest
index d7864d9f40e1..aa6422f563b1 100644
--- a/sci-ml/caffe2/Manifest
+++ b/sci-ml/caffe2/Manifest
@@ -12,7 +12,7 @@ DIST caffe2-patches-20240809.tar.gz 15242 BLAKE2B 77503c61487e7d85cca5afcab9a6e6
 DIST pytorch-2.4.1.tar.gz 115029469 BLAKE2B c2909ff27d527bc57cba56b780d3b8cd07a043ab045caa6c6b27857a16f9ad10aaab2116b26226b1e46ee08ffb44007965d914464418e4ae14ca48c3f3f383bb SHA512 7e9b4485e242eaf0d648765c6621d73d95e7107b766646a098175436d1ab2e2b864badd0757a3bab6b7c318233f2120bad9ac07b39bb9e357897919580c87631
 DIST pytorch-2.5.1.tar.gz 116091366 BLAKE2B 7838b17562b94ffc7d798031348689db607dd5eae2a3c35be365972e2b52a2c1b12067068d5aca5ab00cf0977d9c2c3c9ae5337d69534c864c732e6256cbeef6 SHA512 a913a466324a65fa3d79c5e9ad4d605fc7976f0134fda2f81aaa3cea29d56926604999b8a238759646d211e63b47bbb446cdffa86ca8defd8159f11e30301289
 DIST pytorch-2.6.0.tar.gz 119594438 BLAKE2B 3152eb341cf42295e147e59625beb9c06608aa4b78f9618c1c0024b10c1c767715d07fe8c4be52d029ac47f808cd0d5e65c9530ec90d951a64b993083b4067ad SHA512 a70da80ff09d226085e18228132cf6bb236ad8cc47eed52375d0d2a615f09dd33849da947270b5670c184eab60cb8e2adf11d801babfbda7aa621400501d07b0
-EBUILD caffe2-2.4.1-r5.ebuild 8551 BLAKE2B 5d43307717b3af3aa199b8b74403b23b2f7ff7f7f867a7344cbb0b15b6b692020b1b7d51d41e1d4d6d76eed9b71b5c0b18eed6f3f397dd19cbcf8a98a39e5659 SHA512 27697d1162ba83b0db5ad8149071607d10234a38d0fc27aa72eca775a404b6590d81ddfc5fd7d7ebfd96c2115949d5a4323f3679406b23c79c3ebaee50b318f2
-EBUILD caffe2-2.5.1-r9.ebuild 8861 BLAKE2B fc222b4f9a6e25a63b98872ded5a2a9a5ac5adcab20541ba1f7ae1e0c4845c4894710f2f659e4e63f7f29e166a262ea560dbc9bdb6b6383a5dc62cb7799cc0fd SHA512 d15c58fbc48d69466432e9fe42529640c6f6a94a2004da88e708abde011001bb04a2f9253b70c26146c0b2b3f2270b864b099f7c2f2a332a8630c14d0c0c40a2
-EBUILD caffe2-2.6.0-r1.ebuild 8884 BLAKE2B 3c87ff2e81d44f7bf103bb6538f660c50c2a3207eca72a5ce8abd8e046b5c5dddc23a486fc670701df11f0f781e410aa050821ebe6a8e7c961e18fcd923abca3 SHA512 44219a05050808cd7de8601438218868ebe1cc657775bcaecaf12777ec0fa0d4b2c24e9eeea11422346f9319e14f18d7665090f1655be4c20cebfe177c0b1318
+EBUILD caffe2-2.4.1-r6.ebuild 8556 BLAKE2B 77f1809c8ec73b36e16a8c2939bcafb83a995f8074c163e224f0b02230f6c9cde73c2c11aa07b44f4b670b4472c6ba4157262dd14743cad385187bc4145f5af9 SHA512 7747b9ad89f065af36e5d191f24525748a4c6ea020848b0d64f5699e8f9192699ec523708ea6bc52e1973c5792e65330ba7d27e3a0c88d0990094feda9e9a46d
+EBUILD caffe2-2.5.1-r10.ebuild 8866 BLAKE2B bcc3af372f8ab5969903c32bf15cc619d700776133cfa20e723fe126435505e73fcf48985c338c3055b5a206f776fe5744a73441308114356f35e566aae56b28 SHA512 e175f3741c80e0181ec03952dfac7bcfad808275c60d01f18a46073300d575b31f7b161eb715db22a00e8bb13e37c68051e487982a8f020d63da342da24ff56e
+EBUILD caffe2-2.6.0-r2.ebuild 8889 BLAKE2B 0ccf998a778c69efb005967a5ed6110c34ab263e42f5d969fb4bbbed611209ae1c9191df8c47b53c7c194b2612ce86229456dc4fc6383779a8736e4499056bfa SHA512 d5f19c8f8742b0bfacef9656b9ead6bdaf1185360467d6b83cd1bc7e26db2e3e2c925131850a7f2b9bf3b1fadc081e9d6047ae3b86aeacecc689377fe62edca8
 MISC metadata.xml 1228 BLAKE2B 29fa99ccc009ae8ac209cf0a85a8c87992d10301837f44478b4fe60445dbf261cf1c3869cc6e173944b9c98608fcf32a379c766bf7e53267c90fd71e51e1f9a0 SHA512 06d89ac1b62bf215d56ac67408f6027849f3a1948f1a40013aa989f9d879eb9aa9156d9c336cc83d2bb3795271c11d39d34919f749d83dbfbf48cc2d6aa8545e
diff --git a/sci-ml/caffe2/caffe2-2.4.1-r5.ebuild b/sci-ml/caffe2/caffe2-2.4.1-r5.ebuild
deleted file mode 100644
index fef92ded7db4..000000000000
--- a/sci-ml/caffe2/caffe2-2.4.1-r5.ebuild
+++ /dev/null
@@ -1,304 +0,0 @@
-# Copyright 2022-2025 Gentoo Authors
-# Distributed under the terms of the GNU General Public License v2
-
-EAPI=8
-
-PYTHON_COMPAT=( python3_{10..13} )
-ROCM_VERSION=6.1
-inherit python-r1 cmake cuda flag-o-matic prefix rocm toolchain-funcs
-
-MYPN=pytorch
-MYP=${MYPN}-${PV}
-
-DESCRIPTION="A deep learning framework"
-HOMEPAGE="https://pytorch.org/"
-SRC_URI="https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz
-	-> ${MYP}.tar.gz
-	https://dev.gentoo.org/~tupone/distfiles/${PN}-patches-20240809.tar.gz"
-
-S="${WORKDIR}"/${MYP}
-
-LICENSE="BSD"
-SLOT="0"
-KEYWORDS="~amd64"
-IUSE="cuda distributed fbgemm flash gloo mkl mpi nnpack +numpy onednn openblas opencl openmp qnnpack rocm xnnpack"
-RESTRICT="test"
-REQUIRED_USE="
-	${PYTHON_REQUIRED_USE}
-	mpi? ( distributed )
-	gloo? ( distributed )
-	?? ( cuda rocm )
-	rocm? (
-		|| ( ${ROCM_REQUIRED_USE} )
-		!flash
-	)
-"
-
-# CUDA 12 not supported yet: https://github.com/pytorch/pytorch/issues/91122
-RDEPEND="
-	${PYTHON_DEPS}
-	dev-cpp/abseil-cpp:=
-	dev-cpp/gflags:=
-	>=dev-cpp/glog-0.5.0
-	dev-libs/cpuinfo
-	dev-libs/libfmt:=
-	dev-cpp/opentelemetry-cpp
-	dev-libs/protobuf:=
-	dev-libs/pthreadpool
-	dev-libs/sleef[cpu_flags_x86_avx512f(+),cpu_flags_x86_avx(+)]
-	dev-libs/sleef[cpu_flags_x86_sse3(+),cpu_flags_x86_ssse3(+)]
-	dev-libs/sleef[cpu_flags_x86_sse4_1(+),cpu_flags_x86_sse4_2(+)]
-	virtual/lapack
-	sci-ml/onnx
-	sci-ml/foxi
-	cuda? (
-		dev-libs/cudnn
-		>=dev-libs/cudnn-frontend-1.0.3:0/8
-		<dev-util/nvidia-cuda-toolkit-12.5:=[profiler]
-	)
-	fbgemm? ( sci-ml/FBGEMM )
-	gloo? ( sci-ml/gloo[cuda?] )
-	mpi? ( virtual/mpi )
-	nnpack? ( sci-ml/NNPACK )
-	numpy? ( $(python_gen_cond_dep '
-		dev-python/numpy[${PYTHON_USEDEP}]
-		') )
-	onednn? ( sci-ml/oneDNN )
-	opencl? ( virtual/opencl )
-	qnnpack? (
-		!sci-libs/QNNPACK
-		sci-ml/gemmlowp
-	)
-	rocm? (
-		=dev-util/hip-6.1*
-		=dev-libs/rccl-6.1*[${ROCM_USEDEP}]
-		=sci-libs/rocThrust-6.1*[${ROCM_USEDEP}]
-		=sci-libs/rocPRIM-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipBLAS-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipFFT-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipSPARSE-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipRAND-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipCUB-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipSOLVER-6.1*[${ROCM_USEDEP}]
-		=sci-libs/miopen-6.1*[${ROCM_USEDEP}]
-		=dev-util/roctracer-6.1*[${ROCM_USEDEP}]
-
-		=sci-libs/hipBLASLt-6.1*
-		amdgpu_targets_gfx90a? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx90a] )
-		amdgpu_targets_gfx940? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx940] )
-		amdgpu_targets_gfx941? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx941] )
-		amdgpu_targets_gfx942? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx942] )
-	)
-	distributed? (
-		sci-ml/tensorpipe[cuda?]
-		dev-cpp/cpp-httplib
-	)
-	xnnpack? ( sci-ml/XNNPACK )
-	mkl? ( sci-libs/mkl )
-	openblas? ( sci-libs/openblas )
-"
-DEPEND="
-	${RDEPEND}
-	dev-libs/clog
-	dev-libs/psimd
-	dev-libs/FXdiv
-	dev-libs/pocketfft
-	dev-libs/flatbuffers
-	dev-python/pyyaml[${PYTHON_USEDEP}]
-	dev-python/pybind11[${PYTHON_USEDEP}]
-	dev-python/typing-extensions[${PYTHON_USEDEP}]
-	sci-ml/FP16
-	sci-ml/kineto
-	cuda? ( <=dev-libs/cutlass-3.4.1 )
-	onednn? ( sci-ml/ideep )
-"
-
-PATCHES=(
-	../patches/${PN}-2.4.0-gentoo.patch
-	../patches/${PN}-2.4.0-install-dirs.patch
-	../patches/${PN}-1.12.0-glog-0.6.0.patch
-	../patches/${PN}-1.13.1-tensorpipe.patch
-	../patches/${PN}-2.3.0-cudnn_include_fix.patch
-	../patches/${PN}-2.1.2-fix-rpath.patch
-	../patches/${PN}-2.4.0-fix-openmp-link.patch
-	../patches/${PN}-2.4.0-rocm-fix-std-cpp17.patch
-	../patches/${PN}-2.2.2-musl.patch
-	../patches/${PN}-2.4.0-exclude-aotriton.patch
-	../patches/${PN}-2.3.0-fix-rocm-gcc14-clamp.patch
-	../patches/${PN}-2.3.0-fix-libcpp.patch
-	"${FILESDIR}"/${PN}-2.4.0-libfmt-11.patch
-	"${FILESDIR}"/${PN}-2.4.0-cpp-httplib.patch
-	"${FILESDIR}"/${PN}-2.4.0-cstdint.patch
-)
-
-src_prepare() {
-	filter-lto #bug 862672
-	sed -i \
-		-e "/third_party\/gloo/d" \
-		cmake/Dependencies.cmake \
-		|| die
-	cmake_src_prepare
-	pushd torch/csrc/jit/serialization || die
-	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
-	popd
-	# prefixify the hardcoded paths, after all patches are applied
-	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" \
-			-e "s/HIP 1.0/HIP 1.0 REQUIRED/" \
-			-i cmake/public/LoadHIP.cmake || die
-
-		ebegin "HIPifying cuda sources"
-		${EPYTHON} tools/amd_build/build_amd.py || die
-		eend $?
-	fi
-}
-
-src_configure() {
-	if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
-		ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
-		ewarn "These may not be optimal for your GPU."
-		ewarn ""
-		ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
-		ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
-		ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
-		ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
-		ewarn ""
-		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
-		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
-	fi
-
-	local mycmakeargs=(
-		-DBUILD_CUSTOM_PROTOBUF=OFF
-		-DBUILD_SHARED_LIBS=ON
-
-		-DUSE_CCACHE=OFF
-		-DUSE_CUDA=$(usex cuda)
-		-DUSE_DISTRIBUTED=$(usex distributed)
-		-DUSE_MPI=$(usex mpi)
-		-DUSE_FAKELOWP=OFF
-		-DUSE_FBGEMM=$(usex fbgemm)
-		-DUSE_FLASH_ATTENTION=$(usex flash)
-		-DUSE_MEM_EFF_ATTENTION=OFF
-		-DUSE_GFLAGS=ON
-		-DUSE_GLOG=ON
-		-DUSE_GLOO=$(usex gloo)
-		-DUSE_KINETO=OFF # TODO
-		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
-		-DUSE_MKLDNN=$(usex onednn)
-		-DUSE_NNPACK=$(usex nnpack)
-		-DUSE_XNNPACK=$(usex xnnpack)
-		-DUSE_SYSTEM_XNNPACK=$(usex xnnpack)
-		-DUSE_TENSORPIPE=$(usex distributed)
-		-DUSE_PYTORCH_QNNPACK=$(usex qnnpack)
-		-DUSE_NUMPY=$(usex numpy)
-		-DUSE_OPENCL=$(usex opencl)
-		-DUSE_OPENMP=$(usex openmp)
-		-DUSE_ROCM=$(usex rocm)
-		-DUSE_SYSTEM_CPUINFO=ON
-		-DUSE_SYSTEM_PYBIND11=ON
-		-DUSE_UCC=OFF
-		-DUSE_VALGRIND=OFF
-		-DPython_EXECUTABLE="${PYTHON}"
-		-DUSE_ITT=OFF
-		-DUSE_SYSTEM_PTHREADPOOL=ON
-		-DUSE_SYSTEM_PSIMD=ON
-		-DUSE_SYSTEM_FXDIV=ON
-		-DUSE_SYSTEM_FP16=ON
-		-DUSE_SYSTEM_GLOO=ON
-		-DUSE_SYSTEM_ONNX=ON
-		-DUSE_SYSTEM_SLEEF=ON
-		-DUSE_PYTORCH_METAL=OFF
-		-DUSE_XPU=OFF
-
-		-Wno-dev
-		-DTORCH_INSTALL_LIB_DIR="${EPREFIX}"/usr/$(get_libdir)
-		-DLIBSHM_INSTALL_LIB_SUBDIR="${EPREFIX}"/usr/$(get_libdir)
-	)
-
-	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
-		addpredict "/dev/nvidiactl" # bug 867706
-		addpredict "/dev/char"
-		addpredict "/proc/self/task" # bug 926116
-
-		mycmakeargs+=(
-			-DUSE_CUDNN=ON
-			-DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5 7.0}"
-			-DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication Library
-			-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
-		)
-	elif use rocm; then
-		export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
-
-		mycmakeargs+=(
-			-DUSE_NCCL=ON
-			-DUSE_SYSTEM_NCCL=ON
-		)
-
-		# ROCm libraries produce too much warnings
-		append-cxxflags -Wno-deprecated-declarations -Wno-unused-result
-
-		if tc-is-clang; then
-			# fix mangling in LLVM: https://github.com/llvm/llvm-project/issues/85656
-			append-cxxflags -fclang-abi-compat=17
-		fi
-	fi
-
-	if use onednn; then
-		mycmakeargs+=(
-			-DUSE_MKLDNN=ON
-			-DMKLDNN_FOUND=ON
-			-DMKLDNN_LIBRARIES=dnnl
-			-DMKLDNN_INCLUDE_DIR="${ESYSROOT}/usr/include/oneapi/dnnl"
-		)
-	fi
-
-	python_setup
-	cmake_src_configure
-
-	# do not rerun cmake and the build process in src_install
-	sed '/RERUN/,+1d' -i "${BUILD_DIR}"/build.ninja || die
-}
-
-python_install() {
-	python_domodule python/caffe2
-	python_domodule python/torch
-	ln -s ../../../../../include/torch \
-		"${D}$(python_get_sitedir)"/torch/include/torch || die # bug 923269
-}
-
-src_install() {
-	cmake_src_install
-
-	# Used by pytorch ebuild
-	insinto "/var/lib/${PN}"
-	doins "${BUILD_DIR}"/CMakeCache.txt
-
-	rm -rf python
-	mkdir -p python/torch/include || die
-	mv "${ED}"/usr/lib/python*/site-packages/caffe2 python/ || die
-	cp torch/version.py python/torch/ || die
-	python_foreach_impl python_install
-}
diff --git a/sci-ml/caffe2/caffe2-2.4.1-r6.ebuild b/sci-ml/caffe2/caffe2-2.4.1-r6.ebuild
new file mode 100644
index 000000000000..bc95abbd3fca
--- /dev/null
+++ b/sci-ml/caffe2/caffe2-2.4.1-r6.ebuild
@@ -0,0 +1,305 @@
+# Copyright 2022-2025 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+PYTHON_COMPAT=( python3_{10..13} )
+ROCM_VERSION=6.1
+inherit python-single-r1 cmake cuda flag-o-matic prefix rocm toolchain-funcs
+
+MYPN=pytorch
+MYP=${MYPN}-${PV}
+
+DESCRIPTION="A deep learning framework"
+HOMEPAGE="https://pytorch.org/"
+SRC_URI="https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz
+	-> ${MYP}.tar.gz
+	https://dev.gentoo.org/~tupone/distfiles/${PN}-patches-20240809.tar.gz"
+
+S="${WORKDIR}"/${MYP}
+
+LICENSE="BSD"
+SLOT="0"
+KEYWORDS="~amd64"
+IUSE="cuda distributed fbgemm flash gloo mkl mpi nnpack +numpy onednn openblas opencl openmp qnnpack rocm xnnpack"
+RESTRICT="test"
+REQUIRED_USE="
+	${PYTHON_REQUIRED_USE}
+	mpi? ( distributed )
+	gloo? ( distributed )
+	?? ( cuda rocm )
+	rocm? (
+		|| ( ${ROCM_REQUIRED_USE} )
+		!flash
+	)
+"
+
+# CUDA 12 not supported yet: https://github.com/pytorch/pytorch/issues/91122
+RDEPEND="
+	${PYTHON_DEPS}
+	dev-cpp/abseil-cpp:=
+	dev-cpp/gflags:=
+	>=dev-cpp/glog-0.5.0
+	dev-libs/cpuinfo
+	dev-libs/libfmt:=
+	dev-cpp/opentelemetry-cpp
+	dev-libs/protobuf:=
+	dev-libs/pthreadpool
+	dev-libs/sleef[cpu_flags_x86_avx512f(+),cpu_flags_x86_avx(+)]
+	dev-libs/sleef[cpu_flags_x86_sse3(+),cpu_flags_x86_ssse3(+)]
+	dev-libs/sleef[cpu_flags_x86_sse4_1(+),cpu_flags_x86_sse4_2(+)]
+	virtual/lapack
+	sci-ml/onnx
+	sci-ml/foxi
+	cuda? (
+		dev-libs/cudnn
+		>=dev-libs/cudnn-frontend-1.0.3:0/8
+		<dev-util/nvidia-cuda-toolkit-12.5:=[profiler]
+	)
+	fbgemm? ( sci-ml/FBGEMM )
+	gloo? ( sci-ml/gloo[cuda?] )
+	mpi? ( virtual/mpi )
+	nnpack? ( sci-ml/NNPACK )
+	numpy? ( $(python_gen_cond_dep '
+		dev-python/numpy[${PYTHON_USEDEP}]
+		') )
+	onednn? ( sci-ml/oneDNN )
+	opencl? ( virtual/opencl )
+	qnnpack? (
+		!sci-libs/QNNPACK
+		sci-ml/gemmlowp
+	)
+	rocm? (
+		=dev-util/hip-6.1*
+		=dev-libs/rccl-6.1*[${ROCM_USEDEP}]
+		=sci-libs/rocThrust-6.1*[${ROCM_USEDEP}]
+		=sci-libs/rocPRIM-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipBLAS-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipFFT-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipSPARSE-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipRAND-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipCUB-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipSOLVER-6.1*[${ROCM_USEDEP}]
+		=sci-libs/miopen-6.1*[${ROCM_USEDEP}]
+		=dev-util/roctracer-6.1*[${ROCM_USEDEP}]
+
+		=sci-libs/hipBLASLt-6.1*
+		amdgpu_targets_gfx90a? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx90a] )
+		amdgpu_targets_gfx940? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx940] )
+		amdgpu_targets_gfx941? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx941] )
+		amdgpu_targets_gfx942? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx942] )
+	)
+	distributed? (
+		sci-ml/tensorpipe[cuda?]
+		dev-cpp/cpp-httplib
+	)
+	xnnpack? ( sci-ml/XNNPACK )
+	mkl? ( sci-libs/mkl )
+	openblas? ( sci-libs/openblas )
+"
+DEPEND="
+	${RDEPEND}
+	dev-libs/clog
+	dev-libs/psimd
+	dev-libs/FXdiv
+	dev-libs/pocketfft
+	dev-libs/flatbuffers
+	sci-ml/FP16
+	sci-ml/kineto
+	$(python_gen_cond_dep '
+		dev-python/pybind11[${PYTHON_USEDEP}]
+		dev-python/pyyaml[${PYTHON_USEDEP}]
+		dev-python/typing-extensions[${PYTHON_USEDEP}]
+	')
+	cuda? ( <=dev-libs/cutlass-3.4.1 )
+	onednn? ( sci-ml/ideep )
+"
+
+PATCHES=(
+	../patches/${PN}-2.4.0-gentoo.patch
+	../patches/${PN}-2.4.0-install-dirs.patch
+	../patches/${PN}-1.12.0-glog-0.6.0.patch
+	../patches/${PN}-1.13.1-tensorpipe.patch
+	../patches/${PN}-2.3.0-cudnn_include_fix.patch
+	../patches/${PN}-2.1.2-fix-rpath.patch
+	../patches/${PN}-2.4.0-fix-openmp-link.patch
+	../patches/${PN}-2.4.0-rocm-fix-std-cpp17.patch
+	../patches/${PN}-2.2.2-musl.patch
+	../patches/${PN}-2.4.0-exclude-aotriton.patch
+	../patches/${PN}-2.3.0-fix-rocm-gcc14-clamp.patch
+	../patches/${PN}-2.3.0-fix-libcpp.patch
+	"${FILESDIR}"/${PN}-2.4.0-libfmt-11.patch
+	"${FILESDIR}"/${PN}-2.4.0-cpp-httplib.patch
+	"${FILESDIR}"/${PN}-2.4.0-cstdint.patch
+)
+
+src_prepare() {
+	filter-lto #bug 862672
+	sed -i \
+		-e "/third_party\/gloo/d" \
+		cmake/Dependencies.cmake \
+		|| die
+	cmake_src_prepare
+	pushd torch/csrc/jit/serialization || die
+	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
+	popd
+	# prefixify the hardcoded paths, after all patches are applied
+	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" \
+			-e "s/HIP 1.0/HIP 1.0 REQUIRED/" \
+			-i cmake/public/LoadHIP.cmake || die
+
+		ebegin "HIPifying cuda sources"
+		${EPYTHON} tools/amd_build/build_amd.py || die
+		eend $?
+	fi
+}
+
+src_configure() {
+	if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
+		ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
+		ewarn "These may not be optimal for your GPU."
+		ewarn ""
+		ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
+		ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
+		ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
+		ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
+		ewarn ""
+		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
+		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
+	fi
+
+	local mycmakeargs=(
+		-DBUILD_CUSTOM_PROTOBUF=OFF
+		-DBUILD_SHARED_LIBS=ON
+
+		-DUSE_CCACHE=OFF
+		-DUSE_CUDA=$(usex cuda)
+		-DUSE_DISTRIBUTED=$(usex distributed)
+		-DUSE_MPI=$(usex mpi)
+		-DUSE_FAKELOWP=OFF
+		-DUSE_FBGEMM=$(usex fbgemm)
+		-DUSE_FLASH_ATTENTION=$(usex flash)
+		-DUSE_MEM_EFF_ATTENTION=OFF
+		-DUSE_GFLAGS=ON
+		-DUSE_GLOG=ON
+		-DUSE_GLOO=$(usex gloo)
+		-DUSE_KINETO=OFF # TODO
+		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
+		-DUSE_MKLDNN=$(usex onednn)
+		-DUSE_NNPACK=$(usex nnpack)
+		-DUSE_XNNPACK=$(usex xnnpack)
+		-DUSE_SYSTEM_XNNPACK=$(usex xnnpack)
+		-DUSE_TENSORPIPE=$(usex distributed)
+		-DUSE_PYTORCH_QNNPACK=$(usex qnnpack)
+		-DUSE_NUMPY=$(usex numpy)
+		-DUSE_OPENCL=$(usex opencl)
+		-DUSE_OPENMP=$(usex openmp)
+		-DUSE_ROCM=$(usex rocm)
+		-DUSE_SYSTEM_CPUINFO=ON
+		-DUSE_SYSTEM_PYBIND11=ON
+		-DUSE_UCC=OFF
+		-DUSE_VALGRIND=OFF
+		-DPython_EXECUTABLE="${PYTHON}"
+		-DUSE_ITT=OFF
+		-DUSE_SYSTEM_PTHREADPOOL=ON
+		-DUSE_SYSTEM_PSIMD=ON
+		-DUSE_SYSTEM_FXDIV=ON
+		-DUSE_SYSTEM_FP16=ON
+		-DUSE_SYSTEM_GLOO=ON
+		-DUSE_SYSTEM_ONNX=ON
+		-DUSE_SYSTEM_SLEEF=ON
+		-DUSE_PYTORCH_METAL=OFF
+		-DUSE_XPU=OFF
+
+		-Wno-dev
+		-DTORCH_INSTALL_LIB_DIR="${EPREFIX}"/usr/$(get_libdir)
+		-DLIBSHM_INSTALL_LIB_SUBDIR="${EPREFIX}"/usr/$(get_libdir)
+	)
+
+	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
+		addpredict "/dev/nvidiactl" # bug 867706
+		addpredict "/dev/char"
+		addpredict "/proc/self/task" # bug 926116
+
+		mycmakeargs+=(
+			-DUSE_CUDNN=ON
+			-DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5 7.0}"
+			-DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication Library
+			-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
+		)
+	elif use rocm; then
+		export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
+
+		mycmakeargs+=(
+			-DUSE_NCCL=ON
+			-DUSE_SYSTEM_NCCL=ON
+		)
+
+		# ROCm libraries produce too much warnings
+		append-cxxflags -Wno-deprecated-declarations -Wno-unused-result
+
+		if tc-is-clang; then
+			# fix mangling in LLVM: https://github.com/llvm/llvm-project/issues/85656
+			append-cxxflags -fclang-abi-compat=17
+		fi
+	fi
+
+	if use onednn; then
+		mycmakeargs+=(
+			-DUSE_MKLDNN=ON
+			-DMKLDNN_FOUND=ON
+			-DMKLDNN_LIBRARIES=dnnl
+			-DMKLDNN_INCLUDE_DIR="${ESYSROOT}/usr/include/oneapi/dnnl"
+		)
+	fi
+
+	cmake_src_configure
+
+	# do not rerun cmake and the build process in src_install
+	sed '/RERUN/,+1d' -i "${BUILD_DIR}"/build.ninja || die
+}
+
+python_install() {
+	python_domodule python/caffe2
+	python_domodule python/torch
+	ln -s ../../../../../include/torch \
+		"${D}$(python_get_sitedir)"/torch/include/torch || die # bug 923269
+}
+
+src_install() {
+	cmake_src_install
+
+	# Used by pytorch ebuild
+	insinto "/var/lib/${PN}"
+	doins "${BUILD_DIR}"/CMakeCache.txt
+
+	rm -rf python
+	mkdir -p python/torch/include || die
+	mv "${ED}"/usr/lib/python*/site-packages/caffe2 python/ || die
+	cp torch/version.py python/torch/ || die
+	python_install
+}
diff --git a/sci-ml/caffe2/caffe2-2.5.1-r10.ebuild b/sci-ml/caffe2/caffe2-2.5.1-r10.ebuild
new file mode 100644
index 000000000000..02c57f637949
--- /dev/null
+++ b/sci-ml/caffe2/caffe2-2.5.1-r10.ebuild
@@ -0,0 +1,331 @@
+# Copyright 2022-2025 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+PYTHON_COMPAT=( python3_{10..13} )
+ROCM_VERSION=6.1
+inherit python-single-r1 cmake cuda flag-o-matic prefix rocm toolchain-funcs
+
+MYPN=pytorch
+MYP=${MYPN}-${PV}
+
+DESCRIPTION="A deep learning framework"
+HOMEPAGE="https://pytorch.org/"
+SRC_URI="https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz
+	-> ${MYP}.tar.gz"
+
+S="${WORKDIR}"/${MYP}
+
+LICENSE="BSD"
+SLOT="0"
+KEYWORDS="~amd64"
+IUSE="cuda distributed fbgemm flash gloo memefficient mkl mpi nnpack +numpy
+	onednn openblas opencl openmp qnnpack rocm xnnpack"
+RESTRICT="test"
+REQUIRED_USE="
+	${PYTHON_REQUIRED_USE}
+	mpi? ( distributed )
+	gloo? ( distributed )
+	?? ( cuda rocm )
+	rocm? (
+		|| ( ${ROCM_REQUIRED_USE} )
+		!flash
+	)
+"
+
+RDEPEND="
+	${PYTHON_DEPS}
+	dev-cpp/abseil-cpp:=
+	dev-cpp/gflags:=
+	>=dev-cpp/glog-0.5.0
+	dev-cpp/nlohmann_json
+	dev-cpp/opentelemetry-cpp
+	dev-libs/cpuinfo
+	dev-libs/libfmt:=
+	dev-libs/protobuf:=
+	dev-libs/pthreadpool
+	dev-libs/sleef
+	sci-ml/onnx
+	sci-ml/foxi
+	virtual/lapack
+	cuda? (
+		dev-libs/cudnn
+		>=dev-libs/cudnn-frontend-1.0.3:0/8
+		dev-util/nvidia-cuda-toolkit:=[profiler]
+	)
+	fbgemm? ( sci-ml/FBGEMM )
+	gloo? ( sci-ml/gloo[cuda?] )
+	mpi? ( virtual/mpi )
+	nnpack? ( sci-ml/NNPACK )
+	numpy? ( $(python_gen_cond_dep '
+		dev-python/numpy[${PYTHON_USEDEP}]
+		') )
+	onednn? ( =sci-ml/oneDNN-3.5* )
+	opencl? ( virtual/opencl )
+	qnnpack? (
+		!sci-libs/QNNPACK
+		sci-ml/gemmlowp
+	)
+	rocm? (
+		=dev-util/hip-6.1*
+		=dev-libs/rccl-6.1*[${ROCM_USEDEP}]
+		=sci-libs/rocThrust-6.1*[${ROCM_USEDEP}]
+		=sci-libs/rocPRIM-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipBLAS-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipFFT-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipSPARSE-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipRAND-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipCUB-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipSOLVER-6.1*[${ROCM_USEDEP}]
+		=sci-libs/miopen-6.1*[${ROCM_USEDEP}]
+		=dev-util/roctracer-6.1*[${ROCM_USEDEP}]
+
+		=sci-libs/hipBLASLt-6.1*
+		amdgpu_targets_gfx90a? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx90a] )
+		amdgpu_targets_gfx940? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx940] )
+		amdgpu_targets_gfx941? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx941] )
+		amdgpu_targets_gfx942? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx942] )
+	)
+	distributed? (
+		sci-ml/tensorpipe[cuda?]
+		dev-cpp/cpp-httplib
+	)
+	xnnpack? ( ~sci-ml/XNNPACK-2024.02.29 )
+	mkl? ( sci-libs/mkl )
+	openblas? ( sci-libs/openblas )
+"
+
+DEPEND="
+	${RDEPEND}
+	dev-libs/flatbuffers
+	dev-libs/FXdiv
+	dev-libs/pocketfft
+	dev-libs/psimd
+	sci-ml/FP16
+	sci-ml/kineto
+	$(python_gen_cond_dep '
+		dev-python/pybind11[${PYTHON_USEDEP}]
+		dev-python/pyyaml[${PYTHON_USEDEP}]
+		dev-python/typing-extensions[${PYTHON_USEDEP}]
+	')
+	cuda? ( <=dev-libs/cutlass-3.4.1 )
+	onednn? ( sci-ml/ideep )
+	qnnpack? ( dev-libs/clog )
+"
+
+PATCHES=(
+	"${FILESDIR}"/${P}-unbundle_fmt.patch
+	"${FILESDIR}"/${P}-unbundle_kineto.patch
+	"${FILESDIR}"/${P}-cudnn_include_fix.patch
+	"${FILESDIR}"/${P}-gentoo.patch
+	"${FILESDIR}"/${PN}-2.4.0-cpp-httplib.patch
+	"${FILESDIR}"/${P}-glog-0.6.0.patch
+	"${FILESDIR}"/${P}-newfix-functorch-install.patch
+)
+
+src_prepare() {
+	filter-lto #bug 862672
+
+	# Unbundle fmt
+	sed -i \
+		-e 's|::fmt-header-only||' \
+		c10/CMakeLists.txt \
+		cmake/Dependencies.cmake \
+		torch/CMakeLists.txt \
+		|| die
+
+	# Drop third_party from CMake tree
+	sed -i \
+		-e '/add_subdirectory.*third_party/d' \
+		CMakeLists.txt \
+		cmake/Dependencies.cmake \
+		cmake/ProtoBuf.cmake \
+		aten/src/ATen/CMakeLists.txt \
+		|| die
+	# Change libc10* path
+	sed -i \
+		-e "/EXPORT/s|DESTINATION lib)|DESTINATION $(get_libdir))|" \
+		c10/cuda/CMakeLists.txt \
+		c10/CMakeLists.txt \
+		c10/hip/CMakeLists.txt \
+		|| die
+	sed -i \
+		-e '/Using pocketfft in directory:/d' \
+		cmake/Dependencies.cmake \
+		|| die
+
+	cmake_src_prepare
+	pushd torch/csrc/jit/serialization || die
+	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
+	popd
+
+	# prefixify the hardcoded paths, after all patches are applied
+	hprefixify \
+		aten/CMakeLists.txt \
+		caffe2/CMakeLists.txt \
+		cmake/Metal.cmake \
+		cmake/Modules/*.cmake \
+		cmake/Modules_CUDA_fix/FindCUDNN.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
+		cmake/public/LoadHIP.cmake \
+		cmake/public/cuda.cmake \
+		cmake/Dependencies.cmake \
+		torch/CMakeLists.txt \
+		CMakeLists.txt
+
+	if use rocm; then
+		sed -e "s:/opt/rocm:/usr:" \
+			-e "s:lib/cmake:$(get_libdir)/cmake:g" \
+			-e "s/HIP 1.0/HIP 1.0 REQUIRED/" \
+			-i cmake/public/LoadHIP.cmake || die
+
+		ebegin "HIPifying cuda sources"
+		${EPYTHON} tools/amd_build/build_amd.py || die
+		eend $?
+	fi
+}
+
+src_configure() {
+	if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
+		ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
+		ewarn "These may not be optimal for your GPU."
+		ewarn ""
+		ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
+		ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
+		ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
+		ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
+		ewarn ""
+		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
+		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
+	fi
+
+	local mycmakeargs=(
+		-DBUILD_CUSTOM_PROTOBUF=OFF
+		-DLIBSHM_INSTALL_LIB_SUBDIR="${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_FAKELOWP=OFF
+		-DUSE_FBGEMM=$(usex fbgemm)
+		-DUSE_FLASH_ATTENTION=$(usex flash)
+		-DUSE_GFLAGS=ON
+		-DUSE_GLOG=ON
+		-DUSE_GLOO=$(usex gloo)
+		-DUSE_ITT=OFF
+		-DUSE_KINETO=OFF # TODO
+		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
+		-DUSE_MEM_EFF_ATTENTION=$(usex memefficient)
+		-DUSE_MKLDNN=$(usex onednn)
+		-DUSE_MPI=$(usex mpi)
+		-DUSE_NCCL=OFF
+		-DUSE_NNPACK=$(usex nnpack)
+		-DUSE_NUMA=OFF
+		-DUSE_NUMPY=$(usex numpy)
+		-DUSE_OPENCL=$(usex opencl)
+		-DUSE_OPENMP=$(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_ONNX=ON
+		-DUSE_SYSTEM_PSIMD=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)
+		-DUSE_UCC=OFF
+		-DUSE_VALGRIND=OFF
+		-DUSE_XNNPACK=$(usex xnnpack)
+		-DUSE_XPU=OFF
+		-Wno-dev
+	)
+
+	if use mkl; then
+		mycmakeargs+=(-DBLAS=MKL)
+	elif use openblas; then
+		mycmakeargs+=(-DBLAS=OpenBLAS)
+	else
+		mycmakeargs+=(-DBLAS=Generic -DBLAS_LIBRARIES=)
+	fi
+
+	if use cuda; then
+		addpredict "/dev/nvidiactl" # bug 867706
+		addpredict "/dev/char"
+		addpredict "/proc/self/task" # bug 926116
+
+		mycmakeargs+=(
+			-DUSE_CUDNN=ON
+			-DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5 7.0}"
+			-DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication Library
+			-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
+		)
+	elif use rocm; then
+		export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
+
+		mycmakeargs+=(
+			-DUSE_NCCL=ON
+			-DUSE_SYSTEM_NCCL=ON
+		)
+
+		# ROCm libraries produce too much warnings
+		append-cxxflags -Wno-deprecated-declarations -Wno-unused-result
+
+		if tc-is-clang; then
+			# fix mangling in LLVM: https://github.com/llvm/llvm-project/issues/85656
+			append-cxxflags -fclang-abi-compat=17
+		fi
+	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}"/$(python_get_sitedir)/torch/lib/libtorch_global_deps.so || die
+}
+
+src_install() {
+	cmake_src_install
+
+	# Used by pytorch ebuild
+	insinto "/var/lib/${PN}"
+	doins "${BUILD_DIR}"/CMakeCache.txt
+	dostrip -x /var/lib/${PN}/functorch.so
+
+	rm -rf python
+	mkdir -p python/torch || die
+	cp torch/version.py python/torch/ || die
+	python_install
+}
diff --git a/sci-ml/caffe2/caffe2-2.5.1-r9.ebuild b/sci-ml/caffe2/caffe2-2.5.1-r9.ebuild
deleted file mode 100644
index 43cefe46218d..000000000000
--- a/sci-ml/caffe2/caffe2-2.5.1-r9.ebuild
+++ /dev/null
@@ -1,330 +0,0 @@
-# Copyright 2022-2025 Gentoo Authors
-# Distributed under the terms of the GNU General Public License v2
-
-EAPI=8
-
-PYTHON_COMPAT=( python3_{10..13} )
-ROCM_VERSION=6.1
-inherit python-r1 cmake cuda flag-o-matic prefix rocm toolchain-funcs
-
-MYPN=pytorch
-MYP=${MYPN}-${PV}
-
-DESCRIPTION="A deep learning framework"
-HOMEPAGE="https://pytorch.org/"
-SRC_URI="https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz
-	-> ${MYP}.tar.gz"
-
-S="${WORKDIR}"/${MYP}
-
-LICENSE="BSD"
-SLOT="0"
-KEYWORDS="~amd64"
-IUSE="cuda distributed fbgemm flash gloo memefficient mkl mpi nnpack +numpy
-	onednn openblas opencl openmp qnnpack rocm xnnpack"
-RESTRICT="test"
-REQUIRED_USE="
-	${PYTHON_REQUIRED_USE}
-	mpi? ( distributed )
-	gloo? ( distributed )
-	?? ( cuda rocm )
-	rocm? (
-		|| ( ${ROCM_REQUIRED_USE} )
-		!flash
-	)
-"
-
-RDEPEND="
-	${PYTHON_DEPS}
-	dev-cpp/abseil-cpp:=
-	dev-cpp/gflags:=
-	>=dev-cpp/glog-0.5.0
-	dev-cpp/nlohmann_json
-	dev-cpp/opentelemetry-cpp
-	dev-libs/cpuinfo
-	dev-libs/libfmt:=
-	dev-libs/protobuf:=
-	dev-libs/pthreadpool
-	dev-libs/sleef
-	sci-ml/onnx
-	sci-ml/foxi
-	virtual/lapack
-	cuda? (
-		dev-libs/cudnn
-		>=dev-libs/cudnn-frontend-1.0.3:0/8
-		dev-util/nvidia-cuda-toolkit:=[profiler]
-	)
-	fbgemm? ( sci-ml/FBGEMM )
-	gloo? ( sci-ml/gloo[cuda?] )
-	mpi? ( virtual/mpi )
-	nnpack? ( sci-ml/NNPACK )
-	numpy? ( $(python_gen_cond_dep '
-		dev-python/numpy[${PYTHON_USEDEP}]
-		') )
-	onednn? ( =sci-ml/oneDNN-3.5* )
-	opencl? ( virtual/opencl )
-	qnnpack? (
-		!sci-libs/QNNPACK
-		sci-ml/gemmlowp
-	)
-	rocm? (
-		=dev-util/hip-6.1*
-		=dev-libs/rccl-6.1*[${ROCM_USEDEP}]
-		=sci-libs/rocThrust-6.1*[${ROCM_USEDEP}]
-		=sci-libs/rocPRIM-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipBLAS-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipFFT-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipSPARSE-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipRAND-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipCUB-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipSOLVER-6.1*[${ROCM_USEDEP}]
-		=sci-libs/miopen-6.1*[${ROCM_USEDEP}]
-		=dev-util/roctracer-6.1*[${ROCM_USEDEP}]
-
-		=sci-libs/hipBLASLt-6.1*
-		amdgpu_targets_gfx90a? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx90a] )
-		amdgpu_targets_gfx940? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx940] )
-		amdgpu_targets_gfx941? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx941] )
-		amdgpu_targets_gfx942? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx942] )
-	)
-	distributed? (
-		sci-ml/tensorpipe[cuda?]
-		dev-cpp/cpp-httplib
-	)
-	xnnpack? ( ~sci-ml/XNNPACK-2024.02.29 )
-	mkl? ( sci-libs/mkl )
-	openblas? ( sci-libs/openblas )
-"
-
-DEPEND="
-	${RDEPEND}
-	dev-libs/flatbuffers
-	dev-libs/FXdiv
-	dev-libs/pocketfft
-	dev-libs/psimd
-	dev-python/pybind11[${PYTHON_USEDEP}]
-	dev-python/pyyaml[${PYTHON_USEDEP}]
-	dev-python/typing-extensions[${PYTHON_USEDEP}]
-	sci-ml/FP16
-	sci-ml/kineto
-	cuda? ( <=dev-libs/cutlass-3.4.1 )
-	onednn? ( sci-ml/ideep )
-	qnnpack? ( dev-libs/clog )
-"
-
-PATCHES=(
-	"${FILESDIR}"/${P}-unbundle_fmt.patch
-	"${FILESDIR}"/${P}-unbundle_kineto.patch
-	"${FILESDIR}"/${P}-cudnn_include_fix.patch
-	"${FILESDIR}"/${P}-gentoo.patch
-	"${FILESDIR}"/${PN}-2.4.0-cpp-httplib.patch
-	"${FILESDIR}"/${P}-glog-0.6.0.patch
-	"${FILESDIR}"/${P}-newfix-functorch-install.patch
-)
-
-src_prepare() {
-	filter-lto #bug 862672
-
-	# Unbundle fmt
-	sed -i \
-		-e 's|::fmt-header-only||' \
-		c10/CMakeLists.txt \
-		cmake/Dependencies.cmake \
-		torch/CMakeLists.txt \
-		|| die
-
-	# Drop third_party from CMake tree
-	sed -i \
-		-e '/add_subdirectory.*third_party/d' \
-		CMakeLists.txt \
-		cmake/Dependencies.cmake \
-		cmake/ProtoBuf.cmake \
-		aten/src/ATen/CMakeLists.txt \
-		|| die
-	# Change libc10* path
-	sed -i \
-		-e "/EXPORT/s|DESTINATION lib)|DESTINATION $(get_libdir))|" \
-		c10/cuda/CMakeLists.txt \
-		c10/CMakeLists.txt \
-		c10/hip/CMakeLists.txt \
-		|| die
-	sed -i \
-		-e '/Using pocketfft in directory:/d' \
-		cmake/Dependencies.cmake \
-		|| die
-
-	cmake_src_prepare
-	pushd torch/csrc/jit/serialization || die
-	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
-	popd
-
-	# prefixify the hardcoded paths, after all patches are applied
-	hprefixify \
-		aten/CMakeLists.txt \
-		caffe2/CMakeLists.txt \
-		cmake/Metal.cmake \
-		cmake/Modules/*.cmake \
-		cmake/Modules_CUDA_fix/FindCUDNN.cmake \
-		cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
-		cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
-		cmake/public/LoadHIP.cmake \
-		cmake/public/cuda.cmake \
-		cmake/Dependencies.cmake \
-		torch/CMakeLists.txt \
-		CMakeLists.txt
-
-	if use rocm; then
-		sed -e "s:/opt/rocm:/usr:" \
-			-e "s:lib/cmake:$(get_libdir)/cmake:g" \
-			-e "s/HIP 1.0/HIP 1.0 REQUIRED/" \
-			-i cmake/public/LoadHIP.cmake || die
-
-		ebegin "HIPifying cuda sources"
-		${EPYTHON} tools/amd_build/build_amd.py || die
-		eend $?
-	fi
-}
-
-src_configure() {
-	if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
-		ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
-		ewarn "These may not be optimal for your GPU."
-		ewarn ""
-		ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
-		ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
-		ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
-		ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
-		ewarn ""
-		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
-		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
-	fi
-
-	local mycmakeargs=(
-		-DBUILD_CUSTOM_PROTOBUF=OFF
-		-DLIBSHM_INSTALL_LIB_SUBDIR="${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_FAKELOWP=OFF
-		-DUSE_FBGEMM=$(usex fbgemm)
-		-DUSE_FLASH_ATTENTION=$(usex flash)
-		-DUSE_GFLAGS=ON
-		-DUSE_GLOG=ON
-		-DUSE_GLOO=$(usex gloo)
-		-DUSE_ITT=OFF
-		-DUSE_KINETO=OFF # TODO
-		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
-		-DUSE_MEM_EFF_ATTENTION=$(usex memefficient)
-		-DUSE_MKLDNN=$(usex onednn)
-		-DUSE_MPI=$(usex mpi)
-		-DUSE_NCCL=OFF
-		-DUSE_NNPACK=$(usex nnpack)
-		-DUSE_NUMA=OFF
-		-DUSE_NUMPY=$(usex numpy)
-		-DUSE_OPENCL=$(usex opencl)
-		-DUSE_OPENMP=$(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_ONNX=ON
-		-DUSE_SYSTEM_PSIMD=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)
-		-DUSE_UCC=OFF
-		-DUSE_VALGRIND=OFF
-		-DUSE_XNNPACK=$(usex xnnpack)
-		-DUSE_XPU=OFF
-		-Wno-dev
-	)
-
-	if use mkl; then
-		mycmakeargs+=(-DBLAS=MKL)
-	elif use openblas; then
-		mycmakeargs+=(-DBLAS=OpenBLAS)
-	else
-		mycmakeargs+=(-DBLAS=Generic -DBLAS_LIBRARIES=)
-	fi
-
-	if use cuda; then
-		addpredict "/dev/nvidiactl" # bug 867706
-		addpredict "/dev/char"
-		addpredict "/proc/self/task" # bug 926116
-
-		mycmakeargs+=(
-			-DUSE_CUDNN=ON
-			-DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5 7.0}"
-			-DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication Library
-			-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
-		)
-	elif use rocm; then
-		export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
-
-		mycmakeargs+=(
-			-DUSE_NCCL=ON
-			-DUSE_SYSTEM_NCCL=ON
-		)
-
-		# ROCm libraries produce too much warnings
-		append-cxxflags -Wno-deprecated-declarations -Wno-unused-result
-
-		if tc-is-clang; then
-			# fix mangling in LLVM: https://github.com/llvm/llvm-project/issues/85656
-			append-cxxflags -fclang-abi-compat=17
-		fi
-	fi
-
-	if use onednn; then
-		mycmakeargs+=(
-			-DMKLDNN_FOUND=ON
-			-DMKLDNN_LIBRARIES=dnnl
-			-DMKLDNN_INCLUDE_DIR="${ESYSROOT}/usr/include/oneapi/dnnl"
-		)
-	fi
-
-	python_setup
-	cmake_src_configure
-}
-
-src_compile() {
-	PYTORCH_BUILD_VERSION=${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}"/$(python_get_sitedir)/torch/lib/libtorch_global_deps.so || die
-}
-
-src_install() {
-	cmake_src_install
-
-	# Used by pytorch ebuild
-	insinto "/var/lib/${PN}"
-	doins "${BUILD_DIR}"/CMakeCache.txt
-	dostrip -x /var/lib/${PN}/functorch.so
-
-	rm -rf python
-	mkdir -p python/torch || die
-	cp torch/version.py python/torch/ || die
-	python_foreach_impl python_install
-}
diff --git a/sci-ml/caffe2/caffe2-2.6.0-r1.ebuild b/sci-ml/caffe2/caffe2-2.6.0-r1.ebuild
deleted file mode 100644
index f6fef0622eea..000000000000
--- a/sci-ml/caffe2/caffe2-2.6.0-r1.ebuild
+++ /dev/null
@@ -1,330 +0,0 @@
-# Copyright 2022-2025 Gentoo Authors
-# Distributed under the terms of the GNU General Public License v2
-
-EAPI=8
-
-PYTHON_COMPAT=( python3_{10..13} )
-ROCM_VERSION=6.1
-inherit python-r1 cmake cuda flag-o-matic prefix rocm toolchain-funcs
-
-MYPN=pytorch
-MYP=${MYPN}-${PV}
-
-DESCRIPTION="A deep learning framework"
-HOMEPAGE="https://pytorch.org/"
-SRC_URI="https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz
-	-> ${MYP}.tar.gz"
-
-S="${WORKDIR}"/${MYP}
-
-LICENSE="BSD"
-SLOT="0"
-KEYWORDS="~amd64"
-IUSE="cuda distributed fbgemm flash gloo memefficient mkl mpi nnpack +numpy
-	onednn openblas opencl openmp qnnpack rocm xnnpack"
-RESTRICT="test"
-REQUIRED_USE="
-	${PYTHON_REQUIRED_USE}
-	mpi? ( distributed )
-	gloo? ( distributed )
-	?? ( cuda rocm )
-	rocm? (
-		|| ( ${ROCM_REQUIRED_USE} )
-		!flash
-	)
-"
-
-RDEPEND="
-	${PYTHON_DEPS}
-	dev-cpp/abseil-cpp:=
-	dev-cpp/gflags:=
-	>=dev-cpp/glog-0.5.0
-	dev-cpp/nlohmann_json
-	dev-cpp/opentelemetry-cpp
-	dev-libs/cpuinfo
-	dev-libs/libfmt:=
-	dev-libs/protobuf:=
-	dev-libs/pthreadpool
-	dev-libs/sleef
-	virtual/lapack
-	sci-ml/foxi
-	sci-ml/onnx
-	cuda? (
-		dev-libs/cudnn
-		>=dev-libs/cudnn-frontend-1.0.3:0/8
-		dev-util/nvidia-cuda-toolkit:=[profiler]
-	)
-	fbgemm? ( sci-ml/FBGEMM )
-	gloo? ( sci-ml/gloo[cuda?] )
-	mpi? ( virtual/mpi )
-	nnpack? ( sci-ml/NNPACK )
-	numpy? ( $(python_gen_cond_dep '
-		dev-python/numpy[${PYTHON_USEDEP}]
-		') )
-	onednn? ( =sci-ml/oneDNN-3.5* )
-	opencl? ( virtual/opencl )
-	qnnpack? (
-		!sci-libs/QNNPACK
-		sci-ml/gemmlowp
-	)
-	rocm? (
-		=dev-util/hip-6.1*
-		=dev-libs/rccl-6.1*[${ROCM_USEDEP}]
-		=sci-libs/rocThrust-6.1*[${ROCM_USEDEP}]
-		=sci-libs/rocPRIM-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipBLAS-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipFFT-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipSPARSE-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipRAND-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipCUB-6.1*[${ROCM_USEDEP}]
-		=sci-libs/hipSOLVER-6.1*[${ROCM_USEDEP}]
-		=sci-libs/miopen-6.1*[${ROCM_USEDEP}]
-		=dev-util/roctracer-6.1*[${ROCM_USEDEP}]
-
-		=sci-libs/hipBLASLt-6.1*
-		amdgpu_targets_gfx90a? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx90a] )
-		amdgpu_targets_gfx940? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx940] )
-		amdgpu_targets_gfx941? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx941] )
-		amdgpu_targets_gfx942? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx942] )
-	)
-	distributed? (
-		sci-ml/tensorpipe[cuda?]
-		dev-cpp/cpp-httplib
-	)
-	xnnpack? ( sci-ml/XNNPACK )
-	mkl? ( sci-libs/mkl )
-	openblas? ( sci-libs/openblas )
-"
-
-DEPEND="
-	${RDEPEND}
-	dev-libs/flatbuffers
-	dev-libs/FXdiv
-	dev-libs/pocketfft
-	dev-libs/psimd
-	dev-python/pyyaml[${PYTHON_USEDEP}]
-	dev-python/pybind11[${PYTHON_USEDEP}]
-	dev-python/typing-extensions[${PYTHON_USEDEP}]
-	sci-ml/FP16
-	sci-ml/kineto
-	cuda? ( <=dev-libs/cutlass-3.4.1 )
-	onednn? ( sci-ml/ideep )
-	qnnpack? ( dev-libs/clog )
-"
-
-PATCHES=(
-	"${FILESDIR}"/${PN}-2.5.1-unbundle_fmt.patch
-	"${FILESDIR}"/${PN}-2.5.1-unbundle_kineto.patch
-	"${FILESDIR}"/${PN}-2.5.1-cudnn_include_fix.patch
-	"${FILESDIR}"/${P}-gentoo.patch
-	"${FILESDIR}"/${PN}-2.4.0-cpp-httplib.patch
-	"${FILESDIR}"/${PN}-2.5.1-glog-0.6.0.patch
-	"${FILESDIR}"/${PN}-2.5.1-newfix-functorch-install.patch
-)
-
-src_prepare() {
-	filter-lto #bug 862672
-
-	# Unbundle fmt
-	sed -i \
-		-e 's|::fmt-header-only||' \
-		c10/CMakeLists.txt \
-		cmake/Dependencies.cmake \
-		torch/CMakeLists.txt \
-		|| die
-
-	# Drop third_party from CMake tree
-	sed -i \
-		-e '/add_subdirectory.*third_party/d' \
-		CMakeLists.txt \
-		cmake/Dependencies.cmake \
-		cmake/ProtoBuf.cmake \
-		aten/src/ATen/CMakeLists.txt \
-		|| die
-	# Change libc10* path
-	sed -i \
-		-e "/EXPORT/s|DESTINATION lib)|DESTINATION $(get_libdir))|" \
-		c10/cuda/CMakeLists.txt \
-		c10/CMakeLists.txt \
-		c10/hip/CMakeLists.txt \
-		|| die
-	sed -i \
-		-e '/Using pocketfft in directory:/d' \
-		cmake/Dependencies.cmake \
-		|| die
-
-	cmake_src_prepare
-	pushd torch/csrc/jit/serialization || die
-	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
-	popd
-
-	# prefixify the hardcoded paths, after all patches are applied
-	hprefixify \
-		aten/CMakeLists.txt \
-		caffe2/CMakeLists.txt \
-		cmake/Metal.cmake \
-		cmake/Modules/*.cmake \
-		cmake/Modules_CUDA_fix/FindCUDNN.cmake \
-		cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
-		cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
-		cmake/public/LoadHIP.cmake \
-		cmake/public/cuda.cmake \
-		cmake/Dependencies.cmake \
-		torch/CMakeLists.txt \
-		CMakeLists.txt
-
-	if use rocm; then
-		sed -e "s:/opt/rocm:/usr:" \
-			-e "s:lib/cmake:$(get_libdir)/cmake:g" \
-			-e "s/HIP 1.0/HIP 1.0 REQUIRED/" \
-			-i cmake/public/LoadHIP.cmake || die
-
-		ebegin "HIPifying cuda sources"
-		${EPYTHON} tools/amd_build/build_amd.py || die
-		eend $?
-	fi
-}
-
-src_configure() {
-	if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
-		ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
-		ewarn "These may not be optimal for your GPU."
-		ewarn ""
-		ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
-		ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
-		ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
-		ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
-		ewarn ""
-		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
-		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
-	fi
-
-	local mycmakeargs=(
-		-DBUILD_CUSTOM_PROTOBUF=OFF
-		-DLIBSHM_INSTALL_LIB_SUBDIR="${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_FAKELOWP=OFF
-		-DUSE_FBGEMM=$(usex fbgemm)
-		-DUSE_FLASH_ATTENTION=$(usex flash)
-		-DUSE_GFLAGS=ON
-		-DUSE_GLOG=ON
-		-DUSE_GLOO=$(usex gloo)
-		-DUSE_ITT=OFF
-		-DUSE_KINETO=OFF # TODO
-		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
-		-DUSE_MEM_EFF_ATTENTION=$(usex memefficient)
-		-DUSE_MKLDNN=$(usex onednn)
-		-DUSE_MPI=$(usex mpi)
-		-DUSE_NCCL=OFF
-		-DUSE_NNPACK=$(usex nnpack)
-		-DUSE_NUMA=OFF
-		-DUSE_NUMPY=$(usex numpy)
-		-DUSE_OPENCL=$(usex opencl)
-		-DUSE_OPENMP=$(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_ONNX=ON
-		-DUSE_SYSTEM_PSIMD=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)
-		-DUSE_UCC=OFF
-		-DUSE_VALGRIND=OFF
-		-DUSE_XNNPACK=$(usex xnnpack)
-		-DUSE_XPU=OFF
-		-Wno-dev
-	)
-
-	if use mkl; then
-		mycmakeargs+=(-DBLAS=MKL)
-	elif use openblas; then
-		mycmakeargs+=(-DBLAS=OpenBLAS)
-	else
-		mycmakeargs+=(-DBLAS=Generic -DBLAS_LIBRARIES=)
-	fi
-
-	if use cuda; then
-		addpredict "/dev/nvidiactl" # bug 867706
-		addpredict "/dev/char"
-		addpredict "/proc/self/task" # bug 926116
-
-		mycmakeargs+=(
-			-DUSE_CUDNN=ON
-			-DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5 7.0}"
-			-DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication Library
-			-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
-		)
-	elif use rocm; then
-		export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
-
-		mycmakeargs+=(
-			-DUSE_NCCL=ON
-			-DUSE_SYSTEM_NCCL=ON
-		)
-
-		# ROCm libraries produce too much warnings
-		append-cxxflags -Wno-deprecated-declarations -Wno-unused-result
-
-		if tc-is-clang; then
-			# fix mangling in LLVM: https://github.com/llvm/llvm-project/issues/85656
-			append-cxxflags -fclang-abi-compat=17
-		fi
-	fi
-
-	if use onednn; then
-		mycmakeargs+=(
-			-DMKLDNN_FOUND=ON
-			-DMKLDNN_LIBRARIES=dnnl
-			-DMKLDNN_INCLUDE_DIR="${ESYSROOT}/usr/include/oneapi/dnnl"
-		)
-	fi
-
-	python_setup
-	cmake_src_configure
-}
-
-src_compile() {
-	PYTORCH_BUILD_VERSION=${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}"/$(python_get_sitedir)/torch/lib/libtorch_global_deps.so || die
-}
-
-src_install() {
-	cmake_src_install
-
-	# Used by pytorch ebuild
-	insinto "/var/lib/${PN}"
-	doins "${BUILD_DIR}"/CMakeCache.txt
-	dostrip -x /var/lib/${PN}/functorch.so
-
-	rm -rf python
-	mkdir -p python/torch || die
-	cp torch/version.py python/torch/ || die
-	python_foreach_impl python_install
-}
diff --git a/sci-ml/caffe2/caffe2-2.6.0-r2.ebuild b/sci-ml/caffe2/caffe2-2.6.0-r2.ebuild
new file mode 100644
index 000000000000..b9747ad9eed4
--- /dev/null
+++ b/sci-ml/caffe2/caffe2-2.6.0-r2.ebuild
@@ -0,0 +1,331 @@
+# Copyright 2022-2025 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+PYTHON_COMPAT=( python3_{10..13} )
+ROCM_VERSION=6.1
+inherit python-single-r1 cmake cuda flag-o-matic prefix rocm toolchain-funcs
+
+MYPN=pytorch
+MYP=${MYPN}-${PV}
+
+DESCRIPTION="A deep learning framework"
+HOMEPAGE="https://pytorch.org/"
+SRC_URI="https://github.com/pytorch/${MYPN}/archive/refs/tags/v${PV}.tar.gz
+	-> ${MYP}.tar.gz"
+
+S="${WORKDIR}"/${MYP}
+
+LICENSE="BSD"
+SLOT="0"
+KEYWORDS="~amd64"
+IUSE="cuda distributed fbgemm flash gloo memefficient mkl mpi nnpack +numpy
+	onednn openblas opencl openmp qnnpack rocm xnnpack"
+RESTRICT="test"
+REQUIRED_USE="
+	${PYTHON_REQUIRED_USE}
+	mpi? ( distributed )
+	gloo? ( distributed )
+	?? ( cuda rocm )
+	rocm? (
+		|| ( ${ROCM_REQUIRED_USE} )
+		!flash
+	)
+"
+
+RDEPEND="
+	${PYTHON_DEPS}
+	dev-cpp/abseil-cpp:=
+	dev-cpp/gflags:=
+	>=dev-cpp/glog-0.5.0
+	dev-cpp/nlohmann_json
+	dev-cpp/opentelemetry-cpp
+	dev-libs/cpuinfo
+	dev-libs/libfmt:=
+	dev-libs/protobuf:=
+	dev-libs/pthreadpool
+	dev-libs/sleef
+	virtual/lapack
+	sci-ml/foxi
+	sci-ml/onnx
+	cuda? (
+		dev-libs/cudnn
+		>=dev-libs/cudnn-frontend-1.0.3:0/8
+		dev-util/nvidia-cuda-toolkit:=[profiler]
+	)
+	fbgemm? ( sci-ml/FBGEMM )
+	gloo? ( sci-ml/gloo[cuda?] )
+	mpi? ( virtual/mpi )
+	nnpack? ( sci-ml/NNPACK )
+	numpy? ( $(python_gen_cond_dep '
+		dev-python/numpy[${PYTHON_USEDEP}]
+		') )
+	onednn? ( =sci-ml/oneDNN-3.5* )
+	opencl? ( virtual/opencl )
+	qnnpack? (
+		!sci-libs/QNNPACK
+		sci-ml/gemmlowp
+	)
+	rocm? (
+		=dev-util/hip-6.1*
+		=dev-libs/rccl-6.1*[${ROCM_USEDEP}]
+		=sci-libs/rocThrust-6.1*[${ROCM_USEDEP}]
+		=sci-libs/rocPRIM-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipBLAS-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipFFT-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipSPARSE-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipRAND-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipCUB-6.1*[${ROCM_USEDEP}]
+		=sci-libs/hipSOLVER-6.1*[${ROCM_USEDEP}]
+		=sci-libs/miopen-6.1*[${ROCM_USEDEP}]
+		=dev-util/roctracer-6.1*[${ROCM_USEDEP}]
+
+		=sci-libs/hipBLASLt-6.1*
+		amdgpu_targets_gfx90a? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx90a] )
+		amdgpu_targets_gfx940? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx940] )
+		amdgpu_targets_gfx941? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx941] )
+		amdgpu_targets_gfx942? ( =sci-libs/hipBLASLt-6.1*[amdgpu_targets_gfx942] )
+	)
+	distributed? (
+		sci-ml/tensorpipe[cuda?]
+		dev-cpp/cpp-httplib
+	)
+	xnnpack? ( sci-ml/XNNPACK )
+	mkl? ( sci-libs/mkl )
+	openblas? ( sci-libs/openblas )
+"
+
+DEPEND="
+	${RDEPEND}
+	dev-libs/flatbuffers
+	dev-libs/FXdiv
+	dev-libs/pocketfft
+	dev-libs/psimd
+	sci-ml/FP16
+	sci-ml/kineto
+	$(python_gen_cond_dep '
+		dev-python/pybind11[${PYTHON_USEDEP}]
+		dev-python/pyyaml[${PYTHON_USEDEP}]
+		dev-python/typing-extensions[${PYTHON_USEDEP}]
+	')
+	cuda? ( <=dev-libs/cutlass-3.4.1 )
+	onednn? ( sci-ml/ideep )
+	qnnpack? ( dev-libs/clog )
+"
+
+PATCHES=(
+	"${FILESDIR}"/${PN}-2.5.1-unbundle_fmt.patch
+	"${FILESDIR}"/${PN}-2.5.1-unbundle_kineto.patch
+	"${FILESDIR}"/${PN}-2.5.1-cudnn_include_fix.patch
+	"${FILESDIR}"/${P}-gentoo.patch
+	"${FILESDIR}"/${PN}-2.4.0-cpp-httplib.patch
+	"${FILESDIR}"/${PN}-2.5.1-glog-0.6.0.patch
+	"${FILESDIR}"/${PN}-2.5.1-newfix-functorch-install.patch
+)
+
+src_prepare() {
+	filter-lto #bug 862672
+
+	# Unbundle fmt
+	sed -i \
+		-e 's|::fmt-header-only||' \
+		c10/CMakeLists.txt \
+		cmake/Dependencies.cmake \
+		torch/CMakeLists.txt \
+		|| die
+
+	# Drop third_party from CMake tree
+	sed -i \
+		-e '/add_subdirectory.*third_party/d' \
+		CMakeLists.txt \
+		cmake/Dependencies.cmake \
+		cmake/ProtoBuf.cmake \
+		aten/src/ATen/CMakeLists.txt \
+		|| die
+	# Change libc10* path
+	sed -i \
+		-e "/EXPORT/s|DESTINATION lib)|DESTINATION $(get_libdir))|" \
+		c10/cuda/CMakeLists.txt \
+		c10/CMakeLists.txt \
+		c10/hip/CMakeLists.txt \
+		|| die
+	sed -i \
+		-e '/Using pocketfft in directory:/d' \
+		cmake/Dependencies.cmake \
+		|| die
+
+	cmake_src_prepare
+	pushd torch/csrc/jit/serialization || die
+	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
+	popd
+
+	# prefixify the hardcoded paths, after all patches are applied
+	hprefixify \
+		aten/CMakeLists.txt \
+		caffe2/CMakeLists.txt \
+		cmake/Metal.cmake \
+		cmake/Modules/*.cmake \
+		cmake/Modules_CUDA_fix/FindCUDNN.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
+		cmake/public/LoadHIP.cmake \
+		cmake/public/cuda.cmake \
+		cmake/Dependencies.cmake \
+		torch/CMakeLists.txt \
+		CMakeLists.txt
+
+	if use rocm; then
+		sed -e "s:/opt/rocm:/usr:" \
+			-e "s:lib/cmake:$(get_libdir)/cmake:g" \
+			-e "s/HIP 1.0/HIP 1.0 REQUIRED/" \
+			-i cmake/public/LoadHIP.cmake || die
+
+		ebegin "HIPifying cuda sources"
+		${EPYTHON} tools/amd_build/build_amd.py || die
+		eend $?
+	fi
+}
+
+src_configure() {
+	if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
+		ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
+		ewarn "These may not be optimal for your GPU."
+		ewarn ""
+		ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
+		ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
+		ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
+		ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
+		ewarn ""
+		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
+		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
+	fi
+
+	local mycmakeargs=(
+		-DBUILD_CUSTOM_PROTOBUF=OFF
+		-DLIBSHM_INSTALL_LIB_SUBDIR="${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_FAKELOWP=OFF
+		-DUSE_FBGEMM=$(usex fbgemm)
+		-DUSE_FLASH_ATTENTION=$(usex flash)
+		-DUSE_GFLAGS=ON
+		-DUSE_GLOG=ON
+		-DUSE_GLOO=$(usex gloo)
+		-DUSE_ITT=OFF
+		-DUSE_KINETO=OFF # TODO
+		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
+		-DUSE_MEM_EFF_ATTENTION=$(usex memefficient)
+		-DUSE_MKLDNN=$(usex onednn)
+		-DUSE_MPI=$(usex mpi)
+		-DUSE_NCCL=OFF
+		-DUSE_NNPACK=$(usex nnpack)
+		-DUSE_NUMA=OFF
+		-DUSE_NUMPY=$(usex numpy)
+		-DUSE_OPENCL=$(usex opencl)
+		-DUSE_OPENMP=$(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_ONNX=ON
+		-DUSE_SYSTEM_PSIMD=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)
+		-DUSE_UCC=OFF
+		-DUSE_VALGRIND=OFF
+		-DUSE_XNNPACK=$(usex xnnpack)
+		-DUSE_XPU=OFF
+		-Wno-dev
+	)
+
+	if use mkl; then
+		mycmakeargs+=(-DBLAS=MKL)
+	elif use openblas; then
+		mycmakeargs+=(-DBLAS=OpenBLAS)
+	else
+		mycmakeargs+=(-DBLAS=Generic -DBLAS_LIBRARIES=)
+	fi
+
+	if use cuda; then
+		addpredict "/dev/nvidiactl" # bug 867706
+		addpredict "/dev/char"
+		addpredict "/proc/self/task" # bug 926116
+
+		mycmakeargs+=(
+			-DUSE_CUDNN=ON
+			-DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5 7.0}"
+			-DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication Library
+			-DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
+		)
+	elif use rocm; then
+		export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
+
+		mycmakeargs+=(
+			-DUSE_NCCL=ON
+			-DUSE_SYSTEM_NCCL=ON
+		)
+
+		# ROCm libraries produce too much warnings
+		append-cxxflags -Wno-deprecated-declarations -Wno-unused-result
+
+		if tc-is-clang; then
+			# fix mangling in LLVM: https://github.com/llvm/llvm-project/issues/85656
+			append-cxxflags -fclang-abi-compat=17
+		fi
+	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}"/$(python_get_sitedir)/torch/lib/libtorch_global_deps.so || die
+}
+
+src_install() {
+	cmake_src_install
+
+	# Used by pytorch ebuild
+	insinto "/var/lib/${PN}"
+	doins "${BUILD_DIR}"/CMakeCache.txt
+	dostrip -x /var/lib/${PN}/functorch.so
+
+	rm -rf python
+	mkdir -p python/torch || die
+	cp torch/version.py python/torch/ || die
+	python_install
+}
-- 
cgit v1.2.3