TODO: pytorch xpu attempt

2026-04-16 17:31:09 +03:00 · 2026-04-16 17:31:09 +03:00 · c2111878df
commit c2111878df
parent 6628aed3e0
25 changed files with 1589 additions and 0 deletions
--- a/sci-ml/caffe2/caffe2-2.11.0-r3.ebuild
+++ b/sci-ml/caffe2/caffe2-2.11.0-r3.ebuild
@ -0,0 +1,420 @@
+# Copyright 2022-2026 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+PYTHON_COMPAT=( python3_{11..14} )
+ROCM_VERSION=6.1
+inherit python-single-r1 cmake cuda flag-o-matic prefix rocm toolchain-funcs
+
+MYPN=pytorch
+MYP=${MYPN}-${PV}
+
+# caffe2-2.9.0 depends on future version of composable kernel
+# TODO: replace it with DEPEND in the future
+CK_COMMIT=7fe50dc3da2069d6645d9deb8c017a876472a977
+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 -> ${FLASH_P}.gh.tar.gz"
+
+AOTRITON_PV=0.9.2b
+AOTRITON_PN=aotriton
+AOTRITON_P=${AOTRITON_PN}-${AOTRITON_PV}
+AOTRITON_tar=${AOTRITON_P}-manylinux_2_28_x86_64-rocm6.3-shared.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 opencl 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 )
+"
+
+RDEPEND="
+	${PYTHON_DEPS}
+	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 )
+	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 )
+	opencl? ( virtual/opencl )
+	qnnpack? (
+		!sci-libs/QNNPACK
+		sci-ml/gemmlowp
+		dev-libs/pthreadpool
+	)
+	rocm? (
+		nccl? ( >=dev-libs/rccl-6.3:= <dev-libs/rccl-7.3:= )
+		>=dev-util/hip-6.3:=       <dev-util/hip-7.3:=
+		>=dev-util/roctracer-6.3:= <dev-util/roctracer-7.3:=
+		>=sci-libs/hipBLAS-6.3:=   <sci-libs/hipBLAS-7.3:=[rocsolver(+)]
+		>=sci-libs/hipBLASLt-6.3:= <sci-libs/hipBLASLt-7.3:=
+		>=sci-libs/hipFFT-6.3:=    <sci-libs/hipFFT-7.3:=
+		>=sci-libs/hipRAND-6.3:=   <sci-libs/hipRAND-7.3:=
+		>=sci-libs/hipSOLVER-6.3:= <sci-libs/hipSOLVER-7.3:=
+		>=sci-libs/hipSPARSE-6.3:= <sci-libs/hipSPARSE-7.3:=
+		>=sci-libs/miopen-6.3:=    <sci-libs/miopen-7.3:=
+		>=sci-libs/rocBLAS-6.3:=   <sci-libs/rocBLAS-7.3:=
+		>=sci-libs/rocRAND-6.3:=   <sci-libs/rocRAND-7.3:=
+		>=sci-libs/rocSOLVER-6.3:= <sci-libs/rocSOLVER-7.3:=
+		memefficient? ( =sci-libs/aotriton-bin-0.11*:= )
+		distributed? ( >=dev-util/rocm-smi-6.3:= <dev-util/rocm-smi-7.3:= )
+		cusparselt? ( >=sci-libs/hipsparselt-6.3:= <sci-libs/hipsparselt-7.3:= )
+	)
+	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 )
+"
+
+DEPEND="
+	${RDEPEND}
+	dev-cpp/nlohmann_json
+	dev-libs/flatbuffers
+	dev-libs/FXdiv
+	dev-libs/pocketfft
+	dev-libs/psimd
+	sci-ml/FP16
+	$(python_gen_cond_dep '
+		<dev-python/pybind11-3.0.2[${PYTHON_USEDEP}]
+		dev-python/pyyaml[${PYTHON_USEDEP}]
+		dev-python/typing-extensions[${PYTHON_USEDEP}]
+	')
+	cuda? ( >=dev-libs/cutlass-3.9.2[tools(+)] )
+	onednn? ( sci-ml/ideep )
+	rocm? (
+		>=sci-libs/hipCUB-6.3:=    <sci-libs/hipCUB-7.3:=
+		>=sci-libs/rocPRIM-6.3:=   <sci-libs/rocPRIM-7.3:=
+		>=sci-libs/rocThrust-6.3:= <sci-libs/rocThrust-7.3:=
+	)
+	qnnpack? ( dev-libs/clog )
+"
+
+PATCHES=(
+	"${FILESDIR}"/${PN}-2.5.1-unbundle_fmt.patch
+	"${FILESDIR}"/${PN}-2.5.1-unbundle_kineto.patch
+	"${FILESDIR}"/${PN}-2.8.0-unbundle_pocketfft.patch
+	"${FILESDIR}"/${PN}-2.5.1-cudnn_include_fix.patch
+	"${FILESDIR}"/${PN}-2.4.0-cpp-httplib.patch
+	"${FILESDIR}"/${PN}-2.5.1-glog-0.6.0.patch
+	"${FILESDIR}"/${P}-rocm-fix-std-cpp17.patch
+	"${FILESDIR}"/${PN}-2.7.0-glog-0.7.1.patch
+	"${FILESDIR}"/${PN}-2.7.1-aotriton-fixes.patch
+	"${FILESDIR}"/${PN}-2.8.0-rocm-minus-flash.patch
+	"${FILESDIR}"/${PN}-2.9.0-cmake.patch
+	"${FILESDIR}"/${PN}-2.9.0-rocm-distributed-link.patch
+	"${FILESDIR}"/${PN}-2.9.1-torch_cpu.patch
+	"${FILESDIR}"/${PN}-2.10.0-gentoo.patch
+	"${FILESDIR}"/${P}-mimalloc.patch
+	"${FILESDIR}"/${P}-removekineto-pr178960.patch
+)
+
+src_prepare() {
+	if use cuda && ( use flash || use memefficient ); then
+		mv "${WORKDIR}"/${FLASH_P}/* third_party/${FLASH_PN}/ || die
+	fi
+	filter-lto #bug 862672
+
+	# Unbundle fmt
+	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
+
+	# 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 $(get_libdir))|" \
+		c10/cuda/CMakeLists.txt \
+		c10/CMakeLists.txt \
+		c10/hip/CMakeLists.txt \
+		|| die
+
+	# Change libaotriton path
+	sed -i \
+		-e "s|}/lib|}/\${CMAKE_INSTALL_LIBDIR}|g" \
+		-e "/set(__AOTRITON_LIB/s|lib/|\${CMAKE_INSTALL_LIBDIR}/|g" \
+		cmake/External/aotriton.cmake \
+		|| die
+
+	# Noisy warnings from Logging.h
+	sed -i 's/-Wextra-semi//' cmake/public/utils.cmake || die
+
+	cmake_src_prepare
+	pushd torch/csrc/jit/serialization > /dev/null || die
+	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
+	popd > /dev/null || die
+
+	# prefixify the hardcoded paths, after all patches are applied
+	hprefixify \
+		aten/CMakeLists.txt \
+		caffe2/CMakeLists.txt \
+		cmake/Metal.cmake \
+		cmake/Modules/*.cmake \
+		cmake/Modules_CUDA_fix/FindCUDNN.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
+		cmake/public/LoadHIP.cmake \
+		cmake/public/cuda.cmake \
+		cmake/Dependencies.cmake \
+		torch/CMakeLists.txt \
+		CMakeLists.txt
+
+	if use rocm; then
+		sed -e "s:/opt/rocm:/usr:" \
+			-e "s:lib/cmake:$(get_libdir)/cmake:g" \
+			-i cmake/public/LoadHIP.cmake || die
+
+		# TODO: delete, when caffe2 depends on systemwide composable_kernel
+		sed -e "s:third_party/composable_kernel:../composable_kernel-${CK_COMMIT}:g" \
+			-i aten/src/ATen/CMakeLists.txt || die
+
+		# Bug 959808: fix for gfx101x targets
+		pushd "${WORKDIR}/composable_kernel-${CK_COMMIT}" > /dev/null || die
+		eapply "${FILESDIR}"/composable-kernel-7fe50dc-expand-isa.patch
+		popd > /dev/null || die
+
+		if tc-is-clang; then
+			# Systemwide gcc (for absl and at::TensorBase) + hipcc (llvm>=18) need abi-compat=17.
+			# But systemwide clang>=18 + hipcc (>=llvm-18) need opposite!
+			# See also: https://github.com/llvm/llvm-project/issues/102443#issuecomment-2329726287
+			sed -e '/-fclang-abi-compat=17/d' -i cmake/Dependencies.cmake || die
+		fi
+
+		# Workaround for libc++ 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 ${TORCH_CUDA_ARCH_LIST} ]]; then
+		ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
+		ewarn "These may not be optimal for your GPU."
+		ewarn ""
+		ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
+		ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
+		ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
+		ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
+		ewarn ""
+		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
+		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
+	fi
+
+	local mycmakeargs=(
+		-DBUILD_CUSTOM_PROTOBUF=OFF
+		-DBUILD_TEST=OFF
+		-DLIBSHM_INSTALL_LIB_SUBDIR="${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)
+		-DUSE_KLEIDIAI=OFF # TODO
+		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
+		-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_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_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
+		# bug 867706 926116
+		cuda_add_sandbox
+		addpredict "/dev/char/"
+
+		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 \")"
+			-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:-3.5 7.0}}"
+		fi
+
+	elif use rocm; then
+		export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
+
+		if use memefficient; then
+			export AOTRITON_INSTALLED_PREFIX="${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
+		)
+
+		# ROCm libraries produce too much warnings
+		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}"/$(python_get_sitedir)/torch/lib/libtorch_global_deps.so || die
+}
+
+src_install() {
+	cmake_src_install
+
+	# Used by pytorch ebuild
+	insinto "/var/lib/${PN}"
+	doins "${BUILD_DIR}"/CMakeCache.txt
+
+	rm -rf python
+	mkdir -p python/torch || die
+	cp torch/version.py python/torch/ || die
+	python_install
+}
--- a/sci-ml/caffe2/files/caffe2-2.10.0-gentoo.patch
+++ b/sci-ml/caffe2/files/caffe2-2.10.0-gentoo.patch
@ -0,0 +1,220 @@
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -1084,7 +1084,7 @@
+   set(CMAKE_COLOR_DIAGNOSTICS ON)
+ endif()
+ if(NOT MSVC)
+-  string(APPEND CMAKE_CXX_FLAGS " -O2 -fPIC")
+  string(APPEND CMAKE_CXX_FLAGS " -O2")
+ 
+   # This prevents use of `c10::optional`, `c10::nullopt` etc within the codebase
+   string(APPEND CMAKE_CXX_FLAGS " -DC10_NODEPRECATED")
+@@ -1095,7 +1095,6 @@
+   # Details at http://eigen.tuxfamily.org/bz/show_bug.cgi?id=1459
+   string(APPEND CMAKE_CXX_FLAGS " -Wall")
+   string(APPEND CMAKE_CXX_FLAGS " -Wextra")
+-  append_cxx_flag_if_supported("-Werror=return-type" CMAKE_CXX_FLAGS)
+   append_cxx_flag_if_supported("-Werror=non-virtual-dtor" CMAKE_CXX_FLAGS)
+   append_cxx_flag_if_supported("-Werror=braced-scalar-init" CMAKE_CXX_FLAGS)
+   append_cxx_flag_if_supported("-Werror=range-loop-construct" CMAKE_CXX_FLAGS)
+
+@@ -1083,7 +1082,6 @@
+   endif()
+   append_cxx_flag_if_supported("-fno-math-errno" CMAKE_CXX_FLAGS)
+   append_cxx_flag_if_supported("-fno-trapping-math" CMAKE_CXX_FLAGS)
+-  append_cxx_flag_if_supported("-Werror=format" CMAKE_CXX_FLAGS)
+   if(CMAKE_COMPILER_IS_GNUCXX AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 13)
+     append_cxx_flag_if_supported("-Wno-dangling-reference" CMAKE_CXX_FLAGS)
+     append_cxx_flag_if_supported("-Wno-error=dangling-reference" CMAKE_CXX_FLAGS)
+     append_cxx_flag_if_supported("-Wno-error=redundant-move" CMAKE_CXX_FLAGS)
+--- a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+++ b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+@@ -323,7 +323,7 @@
+ set_target_properties(pytorch_qnnpack PROPERTIES PUBLIC_HEADER include/qnnpack_func.h)
+ 
+ # ---[ Configure clog
+-if(NOT TARGET clog)
+if(FALSE)
+   set(CLOG_BUILD_TESTS OFF CACHE BOOL "")
+   set(CLOG_RUNTIME_TYPE "${CPUINFO_RUNTIME_TYPE}" CACHE STRING "")
+   add_subdirectory(
+@@ -335,7 +335,8 @@
+     target_compile_options(clog PRIVATE "-Wno-unused-result")
+   endif()
+ endif()
+-target_link_libraries(pytorch_qnnpack PUBLIC clog)
+find_library(CLOG_LIBRARY NAMES clog REQUIRED)
+target_link_libraries(pytorch_qnnpack PUBLIC ${CLOG_LIBRARY})
+ 
+ # ---[ Configure cpuinfo
+ if(NOT TARGET cpuinfo AND USE_SYSTEM_CPUINFO)
+--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
+@@ -87,7 +87,7 @@ endif()
+ # Note: the folders that are being commented out have not been properly
+ # addressed yet.
+ 
+-if(NOT MSVC AND USE_XNNPACK)
+if(FALSE)
+   if(NOT TARGET fxdiv)
+     set(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
+     set(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
+@@ -1212,7 +1212,6 @@ if(USE_XPU)
+ endif()
+ 
+ if(NOT MSVC AND USE_XNNPACK)
+-  TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)
+ endif()
+ 
+ # ==========================================================
+@@ -1324,17 +1323,6 @@
+ target_include_directories(torch_cpu PRIVATE
+   "/usr/include/kineto")
+ 
+-if(USE_KINETO)
+-  target_include_directories(torch_cpu PRIVATE
+-    ${TORCH_ROOT}/third_party/kineto/libkineto/src)
+-endif()
+-
+-target_include_directories(torch_cpu PRIVATE
+-  ${TORCH_ROOT}/third_party/cpp-httplib)
+-
+-target_include_directories(torch_cpu PRIVATE
+-  ${TORCH_ROOT}/third_party/nlohmann/include)
+-
+ install(DIRECTORY
+   "${TORCH_SRC_DIR}/csrc"
+   "${TORCH_SRC_DIR}/headeronly"
+--- a/cmake/Codegen.cmake
+++ b/cmake/Codegen.cmake
+@@ -64,7 +64,7 @@ if(INTERN_BUILD_ATEN_OPS)
+   if(MSVC)
+     set(OPT_FLAG "/fp:strict ")
+   else(MSVC)
+-    set(OPT_FLAG "-O3 ")
+    set(OPT_FLAG " ")
+     if("${CMAKE_BUILD_TYPE}" MATCHES "Debug")
+       set(OPT_FLAG " ")
+     endif()
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -492,7 +492,9 @@
+       set_property(TARGET pytorch_qnnpack PROPERTY POSITION_INDEPENDENT_CODE ON)
+       set_property(TARGET cpuinfo PROPERTY POSITION_INDEPENDENT_CODE ON)
+       # QNNPACK depends on gemmlowp headers
+-      target_include_directories(pytorch_qnnpack PRIVATE "${CAFFE2_THIRD_PARTY_ROOT}/gemmlowp")
+      find_package(gemmlowp REQUIRED)
+      get_target_property(GEMMLOWP_INCLUDE_DIRS gemmlowp::gemmlowp INTERFACE_INCLUDE_DIRECTORIES)
+      target_include_directories(pytorch_qnnpack PRIVATE ${GEMMLOWP_INCLUDE_DIRS})
+     endif()
+ 
+     list(APPEND Caffe2_DEPENDENCY_LIBS pytorch_qnnpack)
+@@ -593,7 +591,7 @@
+   find_library(microkernels-prod_LIBRARY microkernels-prod)
+   set_property(TARGET XNNPACK PROPERTY IMPORTED_LOCATION "${XNNPACK_LIBRARY}")
+   set_property(TARGET microkernels-prod PROPERTY IMPORTED_LOCATION "${microkernels-prod_LIBRARY}")
+-  if(NOT XNNPACK_LIBRARY OR NOT microkernels-prod_LIBRARY)
+  if(FALSE)
+     message(FATAL_ERROR "Cannot find XNNPACK")
+   endif()
+   message("-- Found XNNPACK: ${XNNPACK_LIBRARY}")
+@@ -674,7 +672,7 @@ if(BUILD_TEST OR BUILD_MOBILE_BENCHMARK OR BUILD_MOBILE_TEST)
+ endif()
+ 
+ # ---[ FBGEMM
+-if(USE_FBGEMM)
+if(FALSE)
+   set(CAFFE2_THIRD_PARTY_ROOT "${PROJECT_SOURCE_DIR}/third_party")
+   if(NOT DEFINED FBGEMM_SOURCE_DIR)
+     set(FBGEMM_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/fbgemm" CACHE STRING "FBGEMM source directory")
+@@ -698,6 +696,7 @@ if(USE_FBGEMM)
+ endif()
+ 
+ if(USE_FBGEMM)
+  list(APPEND Caffe2_DEPENDENCY_LIBS fbgemm)
+   caffe2_update_option(USE_FBGEMM ON)
+ else()
+   caffe2_update_option(USE_FBGEMM OFF)
+@@ -1156,7 +1155,6 @@
+     endif()
+     set(TP_BUILD_LIBUV ON CACHE BOOL "" FORCE)
+     add_compile_options(-DTORCH_USE_LIBUV)
+-    include_directories(BEFORE SYSTEM ${CMAKE_CURRENT_LIST_DIR}/../third_party/tensorpipe/third_party/libuv/include)
+     set(TP_STATIC_OR_SHARED STATIC CACHE STRING "" FORCE)
+ 
+     # Tensorpipe uses cuda_add_library
+@@ -1676,11 +1676,9 @@
+ 
+ # Include cpp-httplib
+ add_library(httplib INTERFACE IMPORTED)
+-target_include_directories(httplib SYSTEM INTERFACE ${PROJECT_SOURCE_DIR}/third_party/cpp-httplib)
+ 
+ # Include nlohmann-json
+ add_library(nlohmann INTERFACE IMPORTED)
+-include_directories(nlohmann SYSTEM INTERFACE ${PROJECT_SOURCE_DIR}/third_party/nlohmann/include)
+ 
+ # Include moodycamel
+ add_library(moodycamel INTERFACE IMPORTED)
+--- a/cmake/External/nnpack.cmake
+++ b/cmake/External/nnpack.cmake
+@@ -56,7 +56,7 @@
+   set(PTHREADPOOL_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/pthreadpool" CACHE STRING "pthreadpool source directory")
+   set(GOOGLETEST_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/googletest" CACHE STRING "Google Test source directory")
+ 
+-  if(NOT TARGET nnpack)
+  if(FALSE)
+     set(NNPACK_BUILD_TESTS OFF CACHE BOOL "")
+     set(NNPACK_BUILD_BENCHMARKS OFF CACHE BOOL "")
+     set(NNPACK_LIBRARY_TYPE "static" CACHE STRING "")
+--- a/aten/src/ATen/CMakeLists.txt	2025-02-27 14:23:02.402742165 +0100
+++ b/aten/src/ATen/CMakeLists.txt	2025-02-27 14:23:40.445850718 +0100
+@@ -448,8 +448,6 @@
+ if(USE_CUDA)
+   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/cuda)
+   # Next two lines are needed because TunableOp uses third-party/fmt
+-  list(APPEND ATen_CUDA_INCLUDE $<TARGET_PROPERTY:fmt::fmt-header-only,INTERFACE_INCLUDE_DIRECTORIES>)
+-  list(APPEND ATen_CUDA_DEPENDENCY_LIBS fmt::fmt-header-only)
+   list(APPEND ATen_CUDA_CU_SRCS
+     ${cuda_cu}
+     ${native_cuda_cu}
+@@ -521,8 +519,6 @@
+   endif()
+
+   # Next two lines are needed because TunableOp uses third-party/fmt
+-  list(APPEND ATen_HIP_INCLUDE $<TARGET_PROPERTY:fmt::fmt-header-only,INTERFACE_INCLUDE_DIRECTORIES>)
+-  list(APPEND ATen_HIP_DEPENDENCY_LIBS fmt::fmt-header-only)
+   if(USE_FLASH_ATTENTION AND USE_ROCM_CK_SDPA)
+     list(APPEND ATen_HIP_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/native/transformers/hip/flash_attn/ck)
+   endif()
+--- a/torch/CMakeLists.txt
+++ b/torch/CMakeLists.txt
+@@ -59,16 +59,10 @@
+     ${CMAKE_BINARY_DIR}/aten/src
+     ${CMAKE_BINARY_DIR}/caffe2/aten/src
+     ${CMAKE_BINARY_DIR}/third_party
+-    ${CMAKE_BINARY_DIR}/third_party/onnx
+ 
+     ${TORCH_ROOT}/third_party/valgrind-headers
+ 
+-    ${TORCH_ROOT}/third_party/gloo
+-    ${TORCH_ROOT}/third_party/onnx
+-    ${TORCH_ROOT}/third_party/flatbuffers/include
+     "/usr/include/kineto"
+-    ${TORCH_ROOT}/third_party/cpp-httplib
+-    ${TORCH_ROOT}/third_party/nlohmann/include
+ 
+     ${TORCH_SRC_DIR}/csrc
+     ${TORCH_SRC_DIR}/csrc/api/include
+--- a/cmake/FlatBuffers.cmake
+++ b/cmake/FlatBuffers.cmake
+@@ -1,10 +1 @@
+-set(FlatBuffers_Include ${PROJECT_SOURCE_DIR}/third_party/flatbuffers/include)
+-file(GLOB FlatBuffers_Library_SRCS
+-  ${FlatBuffers_Include}/flatbuffers/*.h
+-)
+ add_library(flatbuffers INTERFACE)
+-target_sources(
+-  flatbuffers
+-  INTERFACE ${FlatBuffers_Library_SRCS}
+-)
+-target_include_directories(flatbuffers INTERFACE ${FlatBuffers_Include})
--- a/sci-ml/caffe2/files/caffe2-2.10.0-mimalloc.patch
+++ b/sci-ml/caffe2/files/caffe2-2.10.0-mimalloc.patch
@ -0,0 +1,24 @@
+--- a/c10/CMakeLists.txt	2026-02-01 09:22:55.398151644 +0100
+++ b/c10/CMakeLists.txt	2026-02-01 09:30:46.720495230 +0100
+@@ -120,8 +120,9 @@
+   endif()
+ 
+   if(USE_MIMALLOC)
+-    target_link_libraries(c10 PRIVATE "mimalloc-static")
+-    add_dependencies(c10 mimalloc-static)
+    find_package(mimalloc REQUIRED)
+    target_link_libraries(c10 PRIVATE "mimalloc")
+    add_dependencies(c10 mimalloc)
+   endif()
+ 
+   if(LINUX)
+--- a/CMakeLists.txt	2026-02-03 16:47:46.949580063 +0100
+++ b/CMakeLists.txt	2026-02-03 16:48:57.348716061 +0100
+@@ -404,7 +404,6 @@
+ # on Windows and AArch64.
+ option(USE_MIMALLOC_ON_MKL "Use mimalloc on MKL" OFF)
+ if(WIN32 OR (CPU_AARCH64 AND NOT APPLE))
+-  set(USE_MIMALLOC ON)
+ 
+   # Not enable USE_MIMALLOC_ON_MKL due to it caused issue:
+   # https://github.com/pytorch/pytorch/issues/138994
--- a/sci-ml/caffe2/files/caffe2-2.11.0-mimalloc.patch
+++ b/sci-ml/caffe2/files/caffe2-2.11.0-mimalloc.patch
@ -0,0 +1,24 @@
+--- a/c10/CMakeLists.txt	2026-02-01 09:22:55.398151644 +0100
+++ b/c10/CMakeLists.txt	2026-02-01 09:30:46.720495230 +0100
+@@ -120,8 +120,9 @@
+   endif()
+ 
+   if(USE_MIMALLOC)
+-    target_link_libraries(c10 PRIVATE "mimalloc-static")
+-    add_dependencies(c10 mimalloc-static)
+    find_package(mimalloc REQUIRED)
+    target_link_libraries(c10 PRIVATE "mimalloc")
+    add_dependencies(c10 mimalloc)
+   endif()
+ 
+   if(LINUX)
+--- a/CMakeLists.txt	2026-02-03 16:47:46.949580063 +0100
+++ b/CMakeLists.txt	2026-02-03 16:48:57.348716061 +0100
+@@ -404,7 +404,6 @@
+ # on Windows and AArch64.
+ option(USE_MIMALLOC_ON_MKL "Use mimalloc on MKL" OFF)
+ if(WIN32 OR (CPU_AARCH64 AND NOT APPLE))
+-  set(USE_MIMALLOC ON)
+   # Disable OPT flags for mimalloc on aarch64 as it breaks ARMv8.0
+   # Not enable USE_MIMALLOC_ON_MKL due to it caused issue:
+   # See https://github.com/pytorch/pytorch/issues/174344
--- a/sci-ml/caffe2/files/caffe2-2.11.0-removekineto-pr178960.patch
+++ b/sci-ml/caffe2/files/caffe2-2.11.0-removekineto-pr178960.patch
@ -0,0 +1,170 @@
+diff --git a/torch/csrc/autograd/init.cpp b/torch/csrc/autograd/init.cpp
+index 28c1a43..70d8f7b 100644
+--- a/torch/csrc/autograd/init.cpp
+++ b/torch/csrc/autograd/init.cpp
+@@ -301,12 +301,7 @@ PyObject* THPAutograd_initExtension(PyObject* _unused, PyObject* unused) {
+       .def("privateuse1_elapsed_us", &KinetoEvent::privateuse1ElapsedUs)
+       .def(
+           "is_user_annotation",
+-          [](const KinetoEvent& e) {
+-            return e.activityType() ==
+-                (uint8_t)libkineto::ActivityType::USER_ANNOTATION ||
+-                e.activityType() ==
+-                (uint8_t)libkineto::ActivityType::GPU_USER_ANNOTATION;
+-          })
+          [](const KinetoEvent& e) { return e.isUserAnnotation(); })
+       .def("nbytes", [](const KinetoEvent& e) { return e.nBytes(); })
+       // whether the event is hidden
+       .def(
+diff --git a/torch/csrc/autograd/profiler_kineto.cpp b/torch/csrc/autograd/profiler_kineto.cpp
+index dd4598669827e..85de1c3af923f 100644
+--- a/torch/csrc/autograd/profiler_kineto.cpp
+++ b/torch/csrc/autograd/profiler_kineto.cpp
+@@ -1081,6 +1081,13 @@ int64_t KinetoEvent::privateuse1ElapsedUs() const {
+   return -1;
+ }
+ 
+bool KinetoEvent::isUserAnnotation() const {
+  constexpr uint8_t kUserAnnotation = 1;
+  constexpr uint8_t kGpuUserAnnotation = 2;
+  const auto type = activityType();
+  return type == kUserAnnotation || type == kGpuUserAnnotation;
+}
+
+ void KinetoEvent::getPerfEventCounters(std::vector<uint64_t>& in) const {
+   return result_->visit(c10::overloaded(
+       [&in](const ExtraFields<EventType::TorchOp>& e) -> void {
+diff --git a/torch/csrc/autograd/profiler_kineto.h b/torch/csrc/autograd/profiler_kineto.h
+index 777b8a5851ed5..c3135c61245d4 100644
+--- a/torch/csrc/autograd/profiler_kineto.h
+++ b/torch/csrc/autograd/profiler_kineto.h
+@@ -67,6 +67,7 @@ struct TORCH_API KinetoEvent {
+   bool isPythonFunction() const;
+   int64_t cudaElapsedUs() const;
+   int64_t privateuse1ElapsedUs() const;
+  bool isUserAnnotation() const;
+   void getPerfEventCounters(torch::profiler::perf_counters_t& /*in*/) const;
+   extra_meta_t extraMeta() const;
+   std::string metadataJson() const;
+diff --git a/torch/csrc/profiler/collection.cpp b/torch/csrc/profiler/collection.cpp
+index eb29cb7382127..8ee3a9e7707a8 100644
+--- a/torch/csrc/profiler/collection.cpp
+++ b/torch/csrc/profiler/collection.cpp
+@@ -572,11 +572,13 @@ std::string toString(const ExtraFields<EventType::PyCall>& e) {
+       e.callsite_.funcname_.str());
+ }
+ 
+#ifdef USE_KINETO
+ auto scopeToType(at::RecordScope scope) {
+   return scope == at::RecordScope::USER_SCOPE
+       ? libkineto::ActivityType::USER_ANNOTATION
+       : libkineto::ActivityType::CPU_OP;
+ }
+#endif
+ 
+ int64_t torchOpEndNS(
+     const ExtraFields<EventType::TorchOp>& e,
+@@ -625,6 +627,7 @@ std::string Result::overload_name() const {
+       [](const auto& e) -> std::string { return ""; }));
+ }
+ 
+#ifdef USE_KINETO
+ libkineto::ActivityType Result::kinetoType() const {
+   return visit(c10::overloaded(
+       ATTRIBUTE(TorchOp, scopeToType(e.scope_)),
+@@ -637,6 +640,11 @@ libkineto::ActivityType Result::kinetoType() const {
+       ATTRIBUTE(PythonGC, libkineto::ActivityType::PYTHON_FUNCTION),
+       ATTRIBUTE(Kineto, e.activity_type_)));
+ }
+#else
+libkineto::ActivityType Result::kinetoType() const {
+  return libkineto::ActivityType::NONE;
+}
+#endif
+ 
+ uint64_t Result::correlationID() const {
+   return visit(c10::overloaded(
+diff --git a/torch/csrc/profiler/kineto_shim.cpp b/torch/csrc/profiler/kineto_shim.cpp
+index fa232e1a01016..0ddf62997a07e 100644
+--- a/torch/csrc/profiler/kineto_shim.cpp
+++ b/torch/csrc/profiler/kineto_shim.cpp
+@@ -138,6 +138,7 @@ TraceWrapper::TraceWrapper(const int64_t start_time, const std::string& name)
+ }
+ #endif // USE_KINETO
+ 
+#ifdef USE_KINETO
+ activity_t* TraceWrapper::addCPUActivity(
+     const std::string& name,
+     const libkineto::ActivityType type,
+@@ -145,7 +146,6 @@ activity_t* TraceWrapper::addCPUActivity(
+     const uint64_t correlation_id,
+     const int64_t start_time,
+     const int64_t end_time) {
+-#ifdef USE_KINETO
+   TORCH_CHECK((bool)(*this), "Cannot add event to non-existent trace.");
+   cpu_trace_->emplace_activity(cpu_trace_->span, type, name);
+   auto& act = libkineto::CpuTraceBuffer::toRef(cpu_trace_->activities.back());
+@@ -157,10 +157,18 @@ activity_t* TraceWrapper::addCPUActivity(
+     act.endTime = end_time;
+   }
+   return cpu_trace_->activities.back().get();
+}
+ #else
+activity_t* TraceWrapper::addCPUActivity(
+    const std::string& name,
+    const libkineto::ActivityType type,
+    const DeviceAndResource device_and_resource,
+    const uint64_t correlation_id,
+    const int64_t start_time,
+    const int64_t end_time) {
+   return nullptr;
+-#endif // USE_KINETO
+ }
+#endif // USE_KINETO
+ 
+ void TraceWrapper::transferCpuTrace(int64_t end_time) {
+ #ifdef USE_KINETO
+@@ -473,6 +481,7 @@ void logInvariantViolation(
+ 
+ namespace autograd::profiler {
+ c10::DeviceType deviceTypeFromActivity(libkineto::ActivityType activity_type) {
+#ifdef USE_KINETO
+   // PrivateUse1 kineto backend reuse some ActivityTypes,
+   // If PrivateUse1 backend is enabled, this should return
+   // c10::DeviceType::PrivateUse1.
+@@ -524,6 +533,9 @@ c10::DeviceType deviceTypeFromActivity(libkineto::ActivityType activity_type) {
+       return c10::DeviceType::CPU;
+     }
+   }
+#else
+  return c10::DeviceType::CPU;
+#endif // USE_KINETO
+ }
+ 
+ void addMetadataJson(const std::string& key, const std::string& value) {
+diff --git a/torch/csrc/profiler/kineto_shim.h b/torch/csrc/profiler/kineto_shim.h
+index 4f9bdc6770507..ddc00c03fa9cc 100644
+--- a/torch/csrc/profiler/kineto_shim.h
+++ b/torch/csrc/profiler/kineto_shim.h
+@@ -12,7 +12,21 @@
+ #undef USE_KINETO
+ #endif
+ 
+#ifdef USE_KINETO
+ #include <ActivityType.h>
+#else
+namespace libkineto {
+enum class ActivityType : uint8_t {
+  CPU_OP = 0,
+  USER_ANNOTATION,
+  GPU_USER_ANNOTATION,
+  NONE = CPU_OP,
+};
+inline const char* toString(ActivityType) {
+  return "CPU_OP";
+}
+} // namespace libkineto
+#endif
+ 
+ #include <torch/csrc/Export.h>
+ #include <torch/csrc/profiler/api.h>
--- a/sci-ml/caffe2/files/caffe2-2.11.0-rocm-fix-std-cpp17.patch
+++ b/sci-ml/caffe2/files/caffe2-2.11.0-rocm-fix-std-cpp17.patch
@ -0,0 +1,50 @@
+Fix for CXX=clang USE=rocm error: invalid argument '-std=c++17' not allowed with 'C'
+https://github.com/pytorch/pytorch/issues/103222
+--- a/c10/hip/CMakeLists.txt
+++ b/c10/hip/CMakeLists.txt
+@@ -37,6 +37,7 @@ if(NOT BUILD_LIBTORCHLESS)
+ 
+   # Propagate HIP_CXX_FLAGS that were set from Dependencies.cmake
+   target_compile_options(c10_hip PRIVATE ${HIP_CXX_FLAGS})
+  set_target_properties(c10_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+ 
+   # caffe2_hip adds a bunch of dependencies like rocsparse, but c10/hip is supposed to be
+   # minimal.  I'm not sure if we need hip_hcc or not; for now leave it out
+--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
+@@ -1774,6 +1774,7 @@ if(USE_ROCM)
+ 
+   # Since PyTorch files contain HIP headers, these flags are required for the necessary definitions to be added.
+   target_compile_options(torch_hip PUBLIC ${HIP_CXX_FLAGS})  # experiment
+  set_target_properties(torch_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+ 
+   target_link_libraries(torch_hip PUBLIC c10_hip)
+ 
+@@ -2003,6 +2004,7 @@ if(BUILD_TEST)
+       target_include_directories(${test_name} PRIVATE $<INSTALL_INTERFACE:include>)
+       target_include_directories(${test_name} PRIVATE ${Caffe2_CPU_INCLUDE} ${Caffe2_HIP_INCLUDE})
+       target_compile_options(${test_name} PRIVATE ${HIP_CXX_FLAGS})
+      set_target_properties(${test_name} PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+       add_test(NAME ${test_name} COMMAND $<TARGET_FILE:${test_name}>)
+       if(INSTALL_TEST)
+         set_target_properties(${test_name} PROPERTIES INSTALL_RPATH "${CMAKE_INSTALL_RPATH}:${_rpath_portable_origin}/../lib")
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -1014,7 +1014,6 @@
+       list(APPEND HIP_CXX_FLAGS -DUSE_ROCM_CK_GEMM)
+     endif()
+     list(APPEND HIP_HIPCC_FLAGS --offload-compress)
+-    list(APPEND HIP_HIPCC_FLAGS -std=c++17)
+     # Pass device library path for theRock nightly builds
+     if(DEFINED ENV{HIP_DEVICE_LIB_PATH})
+       file(TO_CMAKE_PATH "$ENV{HIP_DEVICE_LIB_PATH}" _hip_device_lib_path)
+--- a/cmake/public/utils.cmake
+++ b/cmake/public/utils.cmake
+@@ -254,6 +254,7 @@ function(caffe2_hip_binary_target target_name_or_src)
+   caffe2_binary_target(${target_name_or_src})
+ 
+   target_compile_options(${__target} PRIVATE ${HIP_CXX_FLAGS})
+  set_target_properties(${__target} PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+   target_include_directories(${__target} PRIVATE ${Caffe2_HIP_INCLUDE})
+ endfunction()
+ 
--- a/sci-ml/caffe2/files/caffe2-2.4.0-cpp-httplib.patch
+++ b/sci-ml/caffe2/files/caffe2-2.4.0-cpp-httplib.patch
@ -0,0 +1,13 @@
+--- a/torch/lib/libshm/CMakeLists.txt	2024-09-04 06:09:51.943752841 +0200
+++ b/torch/lib/libshm/CMakeLists.txt	2024-09-04 06:10:52.243821438 +0200
+@@ -24,6 +24,10 @@
+   CXX_STANDARD 17)
+ target_link_libraries(shm PRIVATE ${TORCH_CPU_LIB})
+ 
+if (USE_DISTRIBUTED)
+  target_link_libraries(shm PRIVATE cpp-httplib)
+endif()
+
+ if(UNIX AND NOT APPLE)
+   include(CheckLibraryExists)
+   find_package(Threads REQUIRED)
--- a/sci-ml/caffe2/files/caffe2-2.5.1-cudnn_include_fix.patch
+++ b/sci-ml/caffe2/files/caffe2-2.5.1-cudnn_include_fix.patch
@ -0,0 +1,11 @@
+--- a/cmake/Dependencies.cmake	2024-04-29 18:37:34.005639858 +0200
+++ b/cmake/Dependencies.cmake	2024-04-29 18:39:29.126587738 +0200
+@@ -1235,7 +1235,7 @@
+   if(CUDNN_VERSION VERSION_LESS 8.5)
+     message(FATAL_ERROR "PyTorch needs CuDNN-8.5 or above, but found ${CUDNN_VERSION}. Builds are still possible with `USE_CUDNN=0`")
+   endif()
+-  set(CUDNN_FRONTEND_INCLUDE_DIR ${CMAKE_CURRENT_LIST_DIR}/../third_party/cudnn_frontend/include)
+  set(CUDNN_FRONTEND_INCLUDE_DIR /opt/cuda/include)
+   target_include_directories(torch::cudnn INTERFACE ${CUDNN_FRONTEND_INCLUDE_DIR})
+ endif()
+ 
--- a/sci-ml/caffe2/files/caffe2-2.5.1-glog-0.6.0.patch
+++ b/sci-ml/caffe2/files/caffe2-2.5.1-glog-0.6.0.patch
@ -0,0 +1,29 @@
+https://github.com/pytorch/pytorch/issues/58054
+
+--- a/c10/util/Logging.cpp
+++ b/c10/util/Logging.cpp
+@@ -192,23 +192,13 @@
+     google::GLOG_WARNING,
+     "The minimum log level that caffe2 will output.");
+ 
+-// Google glog's api does not have an external function that allows one to check
+-// if glog is initialized or not. It does have an internal function - so we are
+-// declaring it here. This is a hack but has been used by a bunch of others too
+-// (e.g. Torch).
+-namespace google {
+-namespace glog_internal_namespace_ {
+-bool IsGoogleLoggingInitialized();
+-} // namespace glog_internal_namespace_
+-} // namespace google
+-
+ namespace c10 {
+ namespace {
+ 
+ void initGoogleLogging(char const* name) {
+ #if !defined(_MSC_VER)
+   // This trick can only be used on UNIX platforms
+-  if (!::google::glog_internal_namespace_::IsGoogleLoggingInitialized())
+  if (!::google::IsGoogleLoggingInitialized())
+ #endif
+   {
+     ::google::InitGoogleLogging(name);
--- a/sci-ml/caffe2/files/caffe2-2.5.1-unbundle_fmt.patch
+++ b/sci-ml/caffe2/files/caffe2-2.5.1-unbundle_fmt.patch
@ -0,0 +1,10 @@
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -1522,7 +1522,6 @@
+ # CMAKE_CXX_FLAGS in ways that break feature checks. Since we already know
+ # `fmt` is compatible with a superset of the compilers that PyTorch is, it
+ # shouldn't be too bad to just disable the checks.
+-set_target_properties(fmt-header-only PROPERTIES INTERFACE_COMPILE_FEATURES "")
+ 
+ list(APPEND Caffe2_DEPENDENCY_LIBS fmt)
+ set(BUILD_SHARED_LIBS ${TEMP_BUILD_SHARED_LIBS} CACHE BOOL "Build shared libs" FORCE)
--- a/sci-ml/caffe2/files/caffe2-2.5.1-unbundle_kineto.patch
+++ b/sci-ml/caffe2/files/caffe2-2.5.1-unbundle_kineto.patch
@ -0,0 +1,33 @@
+--- a/torch/CMakeLists.txt	2024-11-09 20:50:21.806784527 +0100
+++ b/torch/CMakeLists.txt	2024-11-09 20:52:56.303892418 +0100
+@@ -67,7 +67,7 @@
+     ${TORCH_ROOT}/third_party/gloo
+     ${TORCH_ROOT}/third_party/onnx
+     ${TORCH_ROOT}/third_party/flatbuffers/include
+-    ${TORCH_ROOT}/third_party/kineto/libkineto/include
+    "/usr/include/kineto"
+     ${TORCH_ROOT}/third_party/cpp-httplib
+     ${TORCH_ROOT}/third_party/nlohmann/include
+ 
+--- a/caffe2/CMakeLists.txt	2024-11-09 20:58:37.698085144 +0100
+++ b/caffe2/CMakeLists.txt	2024-11-09 20:59:15.771391643 +0100
+@@ -1185,7 +1185,7 @@
+   ${TORCH_ROOT}/third_party/miniz-2.1.0)
+ 
+ target_include_directories(torch_cpu PRIVATE
+-  ${TORCH_ROOT}/third_party/kineto/libkineto/include)
+  "/usr/include/kineto")
+ 
+ if(USE_KINETO)
+   target_include_directories(torch_cpu PRIVATE
+--- a/cmake/Dependencies.cmake	2025-05-04 15:30:00.268862558 +0200
+++ b/cmake/Dependencies.cmake	2025-05-04 15:30:13.275934233 +0200
+@@ -1711,7 +1711,7 @@
+     endif()
+   endif()
+ 
+-  if(NOT TARGET kineto)
+  if(FALSE)
+     add_subdirectory("${KINETO_SOURCE_DIR}")
+     set_property(TARGET kineto PROPERTY POSITION_INDEPENDENT_CODE ON)
+   endif()
--- a/sci-ml/caffe2/files/caffe2-2.6.0-rocm-fix-std-cpp17.patch
+++ b/sci-ml/caffe2/files/caffe2-2.6.0-rocm-fix-std-cpp17.patch
@ -0,0 +1,50 @@
+Fix for CXX=clang USE=rocm error: invalid argument '-std=c++17' not allowed with 'C'
+https://github.com/pytorch/pytorch/issues/103222
+--- a/c10/hip/CMakeLists.txt
+++ b/c10/hip/CMakeLists.txt
+@@ -36,6 +36,7 @@ if(NOT BUILD_LIBTORCHLESS)
+ 
+   # Propagate HIP_CXX_FLAGS that were set from Dependencies.cmake
+   target_compile_options(c10_hip PRIVATE ${HIP_CXX_FLAGS})
+  set_target_properties(c10_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+ 
+   # caffe2_hip adds a bunch of dependencies like rocsparse, but c10/hip is supposed to be
+   # minimal.  I'm not sure if we need hip_hcc or not; for now leave it out
+--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
+@@ -1684,6 +1684,7 @@ if(USE_ROCM)
+ 
+   # Since PyTorch files contain HIP headers, these flags are required for the necessary definitions to be added.
+   target_compile_options(torch_hip PUBLIC ${HIP_CXX_FLAGS})  # experiment
+  set_target_properties(torch_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+ 
+   target_link_libraries(torch_hip PUBLIC c10_hip)
+ 
+@@ -1886,6 +1887,7 @@ if(BUILD_TEST)
+       target_include_directories(${test_name} PRIVATE $<INSTALL_INTERFACE:include>)
+       target_include_directories(${test_name} PRIVATE ${Caffe2_CPU_INCLUDE} ${Caffe2_HIP_INCLUDE})
+       target_compile_options(${test_name} PRIVATE ${HIP_CXX_FLAGS})
+      set_target_properties(${test_name} PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+       add_test(NAME ${test_name} COMMAND $<TARGET_FILE:${test_name}>)
+       if(INSTALL_TEST)
+         set_target_properties(${test_name} PROPERTIES INSTALL_RPATH "${CMAKE_INSTALL_RPATH}:${_rpath_portable_origin}/../lib")
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -1043,7 +1043,6 @@ if(USE_ROCM)
+     list(APPEND HIP_CXX_FLAGS -Wno-duplicate-decl-specifier)
+     list(APPEND HIP_CXX_FLAGS -DCAFFE2_USE_MIOPEN)
+     list(APPEND HIP_CXX_FLAGS -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_HIP)
+-    list(APPEND HIP_CXX_FLAGS -std=c++17)
+     list(APPEND HIP_CXX_FLAGS -DHIPBLAS_V2)
+     if(HIP_NEW_TYPE_ENUMS)
+       list(APPEND HIP_CXX_FLAGS -DHIP_NEW_TYPE_ENUMS)
+--- a/cmake/public/utils.cmake
+++ b/cmake/public/utils.cmake
+@@ -267,6 +267,7 @@ function(caffe2_hip_binary_target target_name_or_src)
+   caffe2_binary_target(${target_name_or_src})
+ 
+   target_compile_options(${__target} PRIVATE ${HIP_CXX_FLAGS})
+  set_target_properties(${__target} PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS OFF)
+   target_include_directories(${__target} PRIVATE ${Caffe2_HIP_INCLUDE})
+ endfunction()
+ 
--- a/sci-ml/caffe2/files/caffe2-2.7.0-glog-0.7.1.patch
+++ b/sci-ml/caffe2/files/caffe2-2.7.0-glog-0.7.1.patch
@ -0,0 +1,11 @@
+--- a/c10/util/Logging.cpp	2025-06-26 11:47:58.901240691 +0200
+++ b/c10/util/Logging.cpp	2025-06-26 11:51:56.959290731 +0200
+@@ -352,7 +352,7 @@
+ 
+ void ShowLogInfoToStderr() {
+   FLAGS_logtostderr = 1;
+-  FLAGS_minloglevel = std::min(FLAGS_minloglevel, google::GLOG_INFO);
+  FLAGS_minloglevel = std::min(FLAGS_minloglevel, int(google::GLOG_INFO));
+ }
+ } // namespace c10
+ 
--- a/sci-ml/caffe2/files/caffe2-2.7.1-aotriton-fixes.patch
+++ b/sci-ml/caffe2/files/caffe2-2.7.1-aotriton-fixes.patch
@ -0,0 +1,27 @@
+Fix installation with aotriton
+
+Upstream bug: https://github.com/pytorch/pytorch/issues/158109
+--- a/cmake/External/aotriton.cmake
+++ b/cmake/External/aotriton.cmake
+@@ -43,10 +43,6 @@ if(NOT __AOTRITON_INCLUDED)
+ 
+   # Note it is INSTALL"ED"
+   if(DEFINED ENV{AOTRITON_INSTALLED_PREFIX})
+-    install(DIRECTORY
+-            $ENV{AOTRITON_INSTALLED_PREFIX}/${CMAKE_INSTALL_LIBDIR}
+-            $ENV{AOTRITON_INSTALLED_PREFIX}/include
+-            DESTINATION ${__AOTRITON_INSTALL_DIR})
+     set(__AOTRITON_INSTALL_DIR "$ENV{AOTRITON_INSTALLED_PREFIX}")
+     message(STATUS "Using Preinstalled AOTriton at ${__AOTRITON_INSTALL_DIR}")
+   elseif(DEFINED ENV{AOTRITON_INSTALL_FROM_SOURCE})
+--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
+@@ -921,7 +921,7 @@ if(USE_ROCM)
+   set(CUDA_LINK_LIBRARIES_KEYWORD PRIVATE)
+   list(APPEND Caffe2_HIP_SRCS ${GENERATED_CXX_TORCH_CUDA})
+   hip_add_library(torch_hip ${Caffe2_HIP_SRCS})
+-  if(USE_FLASH_ATTENTION)
+  if(USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION)
+     target_link_libraries(torch_hip PRIVATE __caffe2_aotriton)
+   endif()
+   set(CUDA_LINK_LIBRARIES_KEYWORD)
--- a/sci-ml/caffe2/files/caffe2-2.8.0-rocm-minus-flash.patch
+++ b/sci-ml/caffe2/files/caffe2-2.8.0-rocm-minus-flash.patch
@ -0,0 +1,86 @@
+Fix use of undeclared identifier 'CHECK_NOSPARSE_CONTIGUOUS_CUDA' with USE='-flash'
+
+Bug: https://github.com/pytorch/pytorch/issues/160826
+--- a/aten/src/ATen/native/transformers/cuda/attention.cu
+++ b/aten/src/ATen/native/transformers/cuda/attention.cu
+@@ -71,6 +71,7 @@
+ #include <ATen/native/transformers/cuda/sdp_utils.h>
+ #include <ATen/native/transformers/sdp_utils_cpp.h>
+ 
+#include <ATen/native/transformers/flash_api_common.h>
+ #ifdef USE_FLASH_ATTENTION
+ // FlashAttention Specific Imports
+ #include <ATen/native/transformers/cuda/flash_attn/flash_api.h>
+--- a/aten/src/ATen/native/transformers/cuda/attention_backward.cu
+++ b/aten/src/ATen/native/transformers/cuda/attention_backward.cu
+@@ -33,6 +33,7 @@
+ #include <ATen/ops/_scaled_dot_product_flash_attention_backward_native.h>
+ #endif
+ 
+#include <ATen/native/transformers/flash_api_common.h>
+ #ifdef USE_FLASH_ATTENTION
+ // FlashAttention Specific Imports
+ #include <ATen/native/transformers/cuda/flash_attn/flash_api.h>
+--- /dev/null
+++ b/aten/src/ATen/native/transformers/flash_api_common.h
+@@ -0,0 +1,28 @@
+#pragma once
+#include <cstdint>
+#include <limits>
+
+#include <ATen/core/Tensor.h>
+#include <c10/util/Exception.h>
+
+#define CHECK_NOSPARSE_CONTIGUOUS_CUDA(TENSOR)                            \
+  TORCH_CHECK(TENSOR.is_cuda(), #TENSOR " must be a CUDA tensor");     \
+  TORCH_CHECK(!TENSOR.is_sparse(), #TENSOR " must be a dense tensor"); \
+  TORCH_CHECK(TENSOR.is_contiguous());
+
+#define CHECK_NOSPARSE_LASTCONTIGUOUS_CUDA(TENSOR)                        \
+  TORCH_CHECK(TENSOR.is_cuda(), #TENSOR " must be a CUDA tensor");     \
+  TORCH_CHECK(!TENSOR.is_sparse(), #TENSOR " must be a dense tensor"); \
+  TORCH_CHECK(                                                         \
+      TENSOR.stride(-1) == 1, #TENSOR ": last dimension must be contiguous");
+
+#define CHECK_ALIGNED_PTR(PTR, ALIGNMENT) \
+  TORCH_CHECK(                         \
+      uint64_t(PTR) % ALIGNMENT == 0, #PTR " is not correctly aligned")
+
+#define ASSIGN_CHECK_OVERFLOW(A, B)                                    \
+  {                                                                    \
+    A = B;                                                             \
+    TORCH_CHECK(                                                    \
+        B < std::numeric_limits<decltype(A)>::max(), #B " overflows"); \
+  }
+--- a/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.h
+++ b/aten/src/ATen/native/transformers/hip/flash_attn/flash_api.h
+@@ -4,28 +4,7 @@
+ #include <ATen/Context.h>
+ #include <ATen/core/Tensor.h>
+ #include <c10/util/Exception.h>
+-
+-#define CHECK_NOSPARSE_CONTIGUOUS_CUDA(TENSOR)                            \
+-  TORCH_CHECK(TENSOR.is_cuda(), #TENSOR " must be a CUDA tensor");     \
+-  TORCH_CHECK(!TENSOR.is_sparse(), #TENSOR " must be a dense tensor"); \
+-  TORCH_CHECK(TENSOR.is_contiguous());
+-
+-#define CHECK_NOSPARSE_LASTCONTIGUOUS_CUDA(TENSOR)                        \
+-  TORCH_CHECK(TENSOR.is_cuda(), #TENSOR " must be a CUDA tensor");     \
+-  TORCH_CHECK(!TENSOR.is_sparse(), #TENSOR " must be a dense tensor"); \
+-  TORCH_CHECK(                                                         \
+-      TENSOR.stride(-1) == 1, #TENSOR ": last dimension must be contiguous");
+-
+-#define CHECK_ALIGNED_PTR(PTR, ALIGNMENT) \
+-  TORCH_CHECK(                         \
+-      uint64_t(PTR) % ALIGNMENT == 0, #PTR " is not correctly aligned")
+-
+-#define ASSIGN_CHECK_OVERFLOW(A, B)                                    \
+-  {                                                                    \
+-    A = B;                                                             \
+-    TORCH_CHECK(                                                    \
+-        B < std::numeric_limits<decltype(A)>::max(), #B " overflows"); \
+-  }
+#include <ATen/native/transformers/flash_api_common.h>
+ 
+ namespace pytorch_flash {
+ 
--- a/sci-ml/caffe2/files/caffe2-2.8.0-unbundle_pocketfft.patch
+++ b/sci-ml/caffe2/files/caffe2-2.8.0-unbundle_pocketfft.patch
@ -0,0 +1,18 @@
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -276,15 +276,8 @@
+ # --- [ PocketFFT
+ set(AT_POCKETFFT_ENABLED 0)
+ if(NOT AT_MKL_ENABLED)
+-  set(POCKETFFT_INCLUDE_DIR "${Torch_SOURCE_DIR}/third_party/pocketfft/")
+-  if(NOT EXISTS "${POCKETFFT_INCLUDE_DIR}")
+-    message(FATAL_ERROR "pocketfft directory not found, expected ${POCKETFFT_INCLUDE_DIR}")
+-  elseif(NOT EXISTS "${POCKETFFT_INCLUDE_DIR}/pocketfft_hdronly.h")
+-    message(FATAL_ERROR "pocketfft headers not found in ${POCKETFFT_INCLUDE_DIR}")
+-  endif()
+ 
+   set(AT_POCKETFFT_ENABLED 1)
+-  message(STATUS "Using pocketfft in directory: ${POCKETFFT_INCLUDE_DIR}")
+ endif()
+ 
+ # ---[ Dependencies
--- a/sci-ml/caffe2/files/caffe2-2.9.0-cmake.patch
+++ b/sci-ml/caffe2/files/caffe2-2.9.0-cmake.patch
@ -0,0 +1,53 @@
+--- a/.ci/pytorch/test_example_code/CMakeLists.txt	2025-10-24 18:25:10.584717735 +0200
+++ b/.ci/pytorch/test_example_code/CMakeLists.txt	2025-10-24 18:25:17.195639297 +0200
+@@ -1,4 +1,4 @@
+-cmake_minimum_required(VERSION 3.0 FATAL_ERROR)
+cmake_minimum_required(VERSION 3.10 FATAL_ERROR)
+ project(simple-torch-test)
+ 
+ find_package(Torch REQUIRED)
+--- a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt	2025-10-24 18:26:46.165583599 +0200
+++ b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt	2025-10-24 18:27:00.085418425 +0200
+@@ -4,7 +4,7 @@
+ # This source code is licensed under the BSD-style license found in the
+ # LICENSE file in the root directory of this source tree.
+ 
+-cmake_minimum_required(VERSION 3.5 FATAL_ERROR)
+cmake_minimum_required(VERSION 3.10 FATAL_ERROR)
+ 
+ include(GNUInstallDirs)
+ 
+--- a/aten/src/ATen/test/test_install/CMakeLists.txt	2025-10-24 18:27:36.180990366 +0200
+++ b/aten/src/ATen/test/test_install/CMakeLists.txt	2025-10-24 18:27:46.812864333 +0200
+@@ -1,4 +1,4 @@
+-cmake_minimum_required(VERSION 3.5)
+cmake_minimum_required(VERSION 3.10)
+ find_package(ATen REQUIRED)
+ include_directories(${ATEN_INCLUDE_DIR})
+ 
+--- a/aten/src/ATen/nnapi/CMakeLists.txt	2025-10-24 18:28:58.948009238 +0200
+++ b/aten/src/ATen/nnapi/CMakeLists.txt	2025-10-24 18:29:11.040865888 +0200
+@@ -1,6 +1,6 @@
+ # Define this to build the NNAPI binding out of tree.
+ if(PYTORCH_NNAPI_STANDALONE)
+-  cmake_minimum_required(VERSION 3.5 FATAL_ERROR)
+  cmake_minimum_required(VERSION 3.10 FATAL_ERROR)
+   project(pytorch_nnapi)
+ 
+   set(CMAKE_CXX_STANDARD 14 CACHE STRING "The C++ standard whose features are requested to build this target.")
+--- a/android/pytorch_android_torchvision/CMakeLists.txt	2025-10-24 18:29:31.653621543 +0200
+++ b/android/pytorch_android_torchvision/CMakeLists.txt	2025-10-24 18:29:41.668502824 +0200
+@@ -1,4 +1,4 @@
+-cmake_minimum_required(VERSION 3.5)
+cmake_minimum_required(VERSION 3.10)
+ project(pytorch_vision_jni CXX)
+ set(CMAKE_CXX_STANDARD 17 CACHE STRING "The C++ standard whose features are requested to build this target.")
+ #_cmake_modify_IGNORE set(CMAKE_VERBOSE_MAKEFILE ON)
+--- a/android/pytorch_android/CMakeLists.txt	2025-10-24 18:29:54.079355700 +0200
+++ b/android/pytorch_android/CMakeLists.txt	2025-10-24 18:30:39.793813797 +0200
+@@ -1,4 +1,4 @@
+-cmake_minimum_required(VERSION 3.5)
+cmake_minimum_required(VERSION 3.10)
+ option(BUILD_LITE_INTERPRETER "Master flag to build pytorch_jni_lite" ON)
+ message(
+   STATUS
--- a/sci-ml/caffe2/files/caffe2-2.9.0-rocm-distributed-link.patch
+++ b/sci-ml/caffe2/files/caffe2-2.9.0-rocm-distributed-link.patch
@ -0,0 +1,28 @@
+USE_ROCM=ON USE_DISTRIBUTED=ON: fix *runtime* error: undefined symbol: rsmi_init
+
+Upstream bug: https://github.com/pytorch/pytorch/issues/158725
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -1058,6 +1058,12 @@ if(USE_ROCM)
+       )
+     endif()
+ 
+    if(USE_DISTRIBUTED)
+      list(APPEND Caffe2_PUBLIC_HIP_DEPENDENCY_LIBS
+        rocm_smi64
+      )
+    endif()
+
+     # ---[ Kernel asserts
+     # Kernel asserts is disabled for ROCm by default.
+     # It can be turned on by turning on the env USE_ROCM_KERNEL_ASSERT to the build system.
+--- a/cmake/public/LoadHIP.cmake
+++ b/cmake/public/LoadHIP.cmake
+@@ -170,6 +170,7 @@ if(HIP_FOUND)
+   find_package_and_print_version(rocthrust REQUIRED)
+   find_package_and_print_version(hipsolver REQUIRED)
+   find_package_and_print_version(rocsolver REQUIRED)
+  find_package_and_print_version(rocm_smi REQUIRED)
+   # workaround cmake 4 build issue
+   if(CMAKE_VERSION VERSION_GREATER_EQUAL "4.0.0")
+     message(WARNING "Work around hiprtc cmake failure for cmake >= 4")
--- a/sci-ml/caffe2/files/caffe2-2.9.1-torch_cpu.patch
+++ b/sci-ml/caffe2/files/caffe2-2.9.1-torch_cpu.patch
@ -0,0 +1,15 @@
+--- a/caffe2/CMakeLists.txt	2026-01-19 20:28:45.127614622 +0100
+++ b/caffe2/CMakeLists.txt	2026-01-19 20:32:04.509213652 +0100
+@@ -1495,6 +1495,12 @@
+   endif()
+ endif()
+ 
+if(USE_DISTRIBUTED)
+  if(USE_MPI)
+    find_package(httplib REQUIRED)
+    target_link_libraries(torch_cpu PRIVATE httplib::httplib)
+  endif()
+endif()
+ if(NOT INTERN_BUILD_MOBILE)
+   if(${CAFFE2_LINK_LOCAL_PROTOBUF})
+     target_link_libraries(torch_cpu INTERFACE protobuf::libprotobuf)
--- a/sci-ml/caffe2/files/composable-kernel-7fe50dc-expand-isa.patch
+++ b/sci-ml/caffe2/files/composable-kernel-7fe50dc-expand-isa.patch
@ -0,0 +1,141 @@
+Fix for "undeclared identifier 'CK_BUFFER_RESOURCE_3RD_DWORD'" for AMDGPU_TARGETS="gfx1012".
+Combines of 3 patches from https://github.com/ROCm/composable_kernel/issues/775#issuecomment-2726315348
+
+Bug: https://bugs.gentoo.org/947583
+Bug: https://bugs.gentoo.org/show_bug.cgi?id=959808
+--- a/include/ck/ck.hpp
+++ b/include/ck/ck.hpp
+@@ -78,7 +78,7 @@
+ #define CK_BUFFER_RESOURCE_3RD_DWORD -1
+ #elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx9__)
+ #define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
+-#elif defined(__gfx103__)
+#elif defined(__gfx101__) || defined(__gfx103__)
+ #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
+ #elif defined(__gfx11__) || defined(__gfx12__)
+ #define CK_BUFFER_RESOURCE_3RD_DWORD 0x31004000
+@@ -86,12 +86,12 @@
+ 
+ // FMA instruction
+ #ifndef __HIP_DEVICE_COMPILE__                   // for host code, define nothing
+-#elif defined(__gfx803__) || defined(__gfx900__) // for GPU code
+-#define CK_USE_AMD_V_MAC_F32
+-#elif defined(__gfx906__) || defined(__gfx9__) || defined(__gfx103__) // for GPU code
+#elif defined(__gfx906__) || defined(__gfx9__) || defined(__gfx103__) || defined(__gfx1011__) || defined(__gfx1012__) // for GPU code
+ #define CK_USE_AMD_V_FMAC_F32
+ #define CK_USE_AMD_V_DOT2_F32_F16
+ #define CK_USE_AMD_V_DOT4_I32_I8
+#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx101__) // for GPU code
+#define CK_USE_AMD_V_MAC_F32
+ #elif defined(__gfx11__) || defined(__gfx12__)
+ #define CK_USE_AMD_V_FMAC_F32
+ #define CK_USE_AMD_V_DOT2_F32_F16
+--- a/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
+++ b/include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_multiple_d_dl.hpp
+@@ -71,7 +71,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+         const Block2CTileMap block_2_ctile_map)
+ {
+ #if(defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) || \
+-    defined(__gfx103__) || defined(__gfx11__) || defined(__gfx12__))
+    defined(__gfx101__) || defined(__gfx103__) || defined(__gfx11__) || defined(__gfx12__))
+ 
+     const index_t num_blocks_per_batch =
+         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+--- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
+++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_multiple_d_dl.hpp
+@@ -50,7 +50,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+         const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
+         const Block2CTileMap block_2_ctile_map)
+ {
+-#if(defined(__gfx906__) || defined(__gfx9__) || defined(__gfx103__) || defined(__gfx11__) || \
+#if(defined(__gfx906__) || defined(__gfx9__) || defined(__gfx101__) || defined(__gfx103__) || defined(__gfx11__) || \
+     defined(__gfx12__))
+ 
+     constexpr index_t shared_block_size =
+--- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_dl.hpp
+@@ -48,7 +48,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+         const Block2CTileMap block_2_ctile_map,
+         const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+ {
+-#if(defined(__gfx906__) || defined(__gfx103__) || defined(__gfx90a__) || defined(__gfx908__) || \
+#if(defined(__gfx906__) || defined(__gfx101__) || defined(__gfx103__) || defined(__gfx90a__) || defined(__gfx908__) || \
+     defined(__gfx94__) || defined(__gfx11__) || defined(__gfx12__))
+     const index_t num_blocks_per_batch =
+         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+--- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_multiple_d_nhwc_kyxc_nhwk.hpp
+@@ -90,7 +90,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+         const Block2CTileMap block_2_ctile_map,
+         const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+ {
+-#if(defined(__gfx906__) || defined(__gfx103__) || defined(__gfx90a__) || defined(__gfx908__) || \
+#if(defined(__gfx906__) || defined(__gfx101__) || defined(__gfx101__) || defined(__gfx103__) || defined(__gfx90a__) || defined(__gfx908__) || \
+     defined(__gfx94__) || defined(__gfx11__) || defined(__gfx12__))
+     // offset base pointer for each work-group
+     const index_t num_blocks_per_batch =
+--- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp
+@@ -106,7 +106,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+         const Block2CTileMap block_2_ctile_map,
+         const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+ {
+-#if(defined(__gfx906__) || defined(__gfx103__) || defined(__gfx11__) || defined(__gfx12__))
+#if(defined(__gfx906__) || defined(__gfx101__) || defined(__gfx103__) || defined(__gfx11__) || defined(__gfx12__))
+     // offset base pointer for each work-group
+     const index_t num_blocks_per_batch =
+         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+--- a/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
+++ b/include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
+@@ -40,7 +40,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+                                       const BElementwiseOperation b_element_op,
+                                       const CDEElementwiseOperation cde_element_op)
+ {
+-#if(defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx103__) || \
+#if(defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx101__) || defined(__gfx101__) || defined(__gfx103__) || \
+     defined(__gfx11__) || defined(__gfx94__) || defined(__gfx12__))
+     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+ 
+--- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_dpp.hpp
+@@ -28,7 +28,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+ #endif
+     kernel_gemm_dpp(const typename GridwiseGemm::Argument karg)
+ {
+-#if(defined(__gfx103__) || defined(__gfx11__))
+#if(defined(__gfx101__) || defined(__gfx103__) || defined(__gfx11__))
+     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+ 
+     const auto a_grid_desc_ak0_m_ak1 = amd_wave_read_first_lane(
+--- a/include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp
+++ b/include/ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp
+@@ -36,7 +36,7 @@ __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+                             const ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch)
+ {
+ #if(defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx94__) || \
+-    defined(__gfx103__) || defined(__gfx11__) || defined(__gfx12__))
+    defined(__gfx101__) || defined(__gfx103__) || defined(__gfx11__) || defined(__gfx12__))
+     GridwiseTensorRearrangeKernel::Run(in_grid_desc,
+                                        p_in_global,
+                                        out_grid_desc,
+--- a/include/ck_tile/core/config.hpp
+++ b/include/ck_tile/core/config.hpp
+@@ -9,6 +9,9 @@
+ #if defined(__gfx942__) || defined(__gfx950__)
+ #define __gfx94__
+ #endif
+#if defined(__gfx1010__) || defined(__gfx1011__) || defined(__gfx1012__)
+#define __gfx101__
+#endif
+ #if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || \
+     defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || \
+     defined(__gfx10_3_generic__)
+@@ -200,7 +203,7 @@
+ #elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || \
+     defined(__gfx9__) // for GPU code
+ #define CK_TILE_BUFFER_RESOURCE_3RD_DWORD 0x00020000
+-#elif defined(__gfx103__) // for GPU code
+#elif defined(__gfx101__) || defined(__gfx103__) // for GPU code
+ #define CK_TILE_BUFFER_RESOURCE_3RD_DWORD 0x31014000
+ #elif defined(__gfx11__) || defined(__gfx12__) // for GPU code
+ #define CK_TILE_BUFFER_RESOURCE_3RD_DWORD 0x31004000
--- a/sci-ml/caffe2/metadata.xml
+++ b/sci-ml/caffe2/metadata.xml
@ -0,0 +1,31 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE pkgmetadata SYSTEM "https://www.gentoo.org/dtd/metadata.dtd">
+<pkgmetadata>
+	<maintainer type="person">
+		<email>tupone@gentoo.org</email>
+		<name>Tupone Alfredo</name>
+	</maintainer>
+	<use>
+		<flag name="cusparselt">Use the CUDA/HIP Sparse Matrix Multiplication</flag>
+		<flag name="distributed">Support distributed applications</flag>
+		<flag name="fbgemm">Use <pkg>sci-ml/FBGEMM</pkg></flag>
+		<flag name="flash">Enable flash attention</flag>
+		<flag name="gloo">Use <pkg>sci-ml/gloo</pkg></flag>
+		<flag name="kineto">Use<pkg>sci-ml/kineto</pkg>profiling library</flag>
+		<flag name="memefficient">Enable mem efficient attention</flag>
+		<flag name="mimalloc">Use <pkg>dev-libs/mimalloc</pkg> as replacement for system malloc</flag>
+		<flag name="mkl">Use <pkg>sci-libs/mkl</pkg> for blas, lapack and sparse blas routines</flag>
+		<flag name="nccl">Use <pkg>dev-libs/rccl</pkg> (NCCL compatible) backend for distributed operations</flag>
+		<flag name="nnpack">Use <pkg>sci-ml/NNPACK</pkg></flag>
+		<flag name="numpy">Add support for math operations through numpy</flag>
+		<flag name="onednn">Use <pkg>sci-ml/oneDNN</pkg></flag>
+		<flag name="openblas">Use <pkg>sci-libs/openblas</pkg> for blas routines</flag>
+		<flag name="openmp">Use OpenMP for parallel code</flag>
+		<flag name="qnnpack">Use QNNPACK</flag>
+		<flag name="rocm">Enable ROCm gpu computing support</flag>
+		<flag name="xnnpack">Use <pkg>sci-ml/XNNPACK</pkg></flag>
+	</use>
+	<upstream>
+		<remote-id type="github">pytorch/pytorch</remote-id>
+	</upstream>
+</pkgmetadata>
--- a/sci-ml/pytorch/files/pytorch-2.10.0-cpp-extension-multilib.patch
+++ b/sci-ml/pytorch/files/pytorch-2.10.0-cpp-extension-multilib.patch
@ -0,0 +1,16 @@
+Fixes multilib (%LIB_DIR% should be replaced in ebuild)
+--- a/torch/utils/cpp_extension.py
+++ b/torch/utils/cpp_extension.py
+@@ -1478,10 +1477,10 @@ def library_paths(device_type: str = "cpu") -> list[str]:
+     paths = [TORCH_LIB_PATH]
+ 
+     if device_type == "cuda" and IS_HIP_EXTENSION:
+-        lib_dir = 'lib'
+        lib_dir = '%LIB_DIR%'
+         paths.append(_join_rocm_home(lib_dir))
+         if HIP_HOME is not None:
+-            paths.append(os.path.join(HIP_HOME, 'lib'))
+            paths.append(os.path.join(HIP_HOME, '%LIB_DIR%'))
+     elif device_type == "cuda":
+         if IS_WINDOWS:
+             lib_dir = os.path.join('lib', 'x64')
--- a/sci-ml/pytorch/files/pytorch-2.9.0-dontbuildagain.patch
+++ b/sci-ml/pytorch/files/pytorch-2.9.0-dontbuildagain.patch
@ -0,0 +1,25 @@
+--- a/setup.py
+++ b/setup.py
+@@ -291,6 +291,7 @@
+ import setuptools.errors
+ from setuptools import Command, Extension, find_packages, setup
+ from setuptools.dist import Distribution
+sys.path[:0] = os.getcwd()
+ 
+ 
+ CWD = Path(__file__).absolute().parent
+@@ -483,14 +483,6 @@
+ 
+     check_submodules()
+     check_pydep("yaml", "pyyaml")
+-    build_pytorch(
+-        version=TORCH_VERSION,
+-        cmake_python_library=CMAKE_PYTHON_LIBRARY.as_posix(),
+-        build_python=not BUILD_LIBTORCH_WHL,
+-        rerun_cmake=RERUN_CMAKE,
+-        cmake_only=CMAKE_ONLY,
+-        cmake=cmake,
+-    )
+ 
+     if CMAKE_ONLY:
+         report(
--- a/sci-ml/pytorch/metadata.xml
+++ b/sci-ml/pytorch/metadata.xml
@ -0,0 +1,14 @@
+<?xml version="1.0" encoding="UTF-8"?>
+<!DOCTYPE pkgmetadata SYSTEM "https://www.gentoo.org/dtd/metadata.dtd">
+<pkgmetadata>
+	<maintainer type="person">
+		<email>tupone@gentoo.org</email>
+		<name>Tupone Alfredo</name>
+	</maintainer>
+	<upstream>
+		<bugs-to>https://github.com/pytorch/pytorch/issues</bugs-to>
+		<changelog>https://github.com/pytorch/pytorch/releases</changelog>
+		<doc>https://pytorch.org/docs/stable/index.html</doc>
+		<remote-id type="github">pytorch/pytorch</remote-id>
+	</upstream>
+</pkgmetadata>
--- a/sci-ml/pytorch/pytorch-2.11.0.ebuild
+++ b/sci-ml/pytorch/pytorch-2.11.0.ebuild
@ -0,0 +1,70 @@
+# Copyright 2022-2026 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+DISTUTILS_USE_PEP517=setuptools
+PYTHON_COMPAT=( python3_{11..14} )
+DISTUTILS_SINGLE_IMPL=1
+DISTUTILS_EXT=1
+inherit distutils-r1 prefix
+
+DESCRIPTION="Tensors and Dynamic neural networks in Python"
+HOMEPAGE="https://pytorch.org/"
+SRC_URI="https://github.com/pytorch/${PN}/archive/refs/tags/v${PV}.tar.gz
+	-> ${P}.tar.gz"
+
+LICENSE="BSD"
+SLOT="0"
+KEYWORDS="~amd64 ~arm64"
+RESTRICT="test"
+
+REQUIRED_USE=${PYTHON_REQUIRED_USE}
+RDEPEND="
+	${PYTHON_DEPS}
+	~sci-ml/caffe2-${PV}[${PYTHON_SINGLE_USEDEP}]
+	$(python_gen_cond_dep '
+		dev-python/sympy[${PYTHON_USEDEP}]
+		dev-python/typing-extensions[${PYTHON_USEDEP}]
+	')
+"
+DEPEND="${RDEPEND}
+	$(python_gen_cond_dep '
+		dev-python/pyyaml[${PYTHON_USEDEP}]
+	')
+"
+
+PATCHES=(
+	"${FILESDIR}"/${PN}-2.9.0-dontbuildagain.patch
+	"${FILESDIR}"/${PN}-2.10.0-cpp-extension-multilib.patch
+)
+
+src_prepare() {
+	# Replace placeholders added by cpp-extension.patch
+	sed -e "s|%LIB_DIR%|$(get_libdir)|g" \
+		-i torch/utils/cpp_extension.py || die
+
+	# Set build dir for pytorch's setup
+	sed -e "/BUILD_DIR/s|build|/var/lib/caffe2/|" \
+		-i tools/setup_helpers/env.py || die
+
+	# Drop legacy from pyproject.toml
+	sed -e "/build-backend/s|:__legacy__||" \
+		-i pyproject.toml || die
+
+	distutils-r1_src_prepare
+
+	hprefixify tools/setup_helpers/env.py
+}
+
+python_compile() {
+	PYTORCH_BUILD_VERSION=${PV} \
+	PYTORCH_BUILD_NUMBER=0 \
+	USE_SYSTEM_LIBS=ON \
+	CMAKE_BUILD_DIR="${BUILD_DIR}" \
+	distutils-r1_python_compile develop sdist
+}
+
+python_install() {
+	USE_SYSTEM_LIBS=ON distutils-r1_python_install
+}