sci-libs/caffe2: add 2.5.1

Signed-off-by: Alfredo Tupone <tupone@gentoo.org>

11 files changed, 559 insertions, 0 deletions

diff --git a/sci-libs/caffe2/Manifest b/sci-libs/caffe2/Manifest
index 1bdb2764edd1..7f0b51661b1b 100644
--- a/sci-libs/caffe2/Manifest
+++ b/sci-libs/caffe2/Manifest
@@ -1,3 +1,4 @@
 DIST caffe2-patches-20240809.tar.gz 15242 BLAKE2B 77503c61487e7d85cca5afcab9a6e638f9833a70861845638cf1b62bc492d7b6650e6db81d53ebb2f39c6313509250d339f725f04d03ec6dd23dd0cf70843d8c SHA512 74b3b0b6671b655ecac93f7436c4ed7cb0157a83aafbf6afcc0811e11cef341cd8f638db1a111bcbb01e1a6dd4daf3a36b96d7a8ce90f04c2fa091bd6e3a142b
 DIST pytorch-2.4.0.tar.gz 115031093 BLAKE2B d206477963977011627df284efa01482fbf57e9fcb5f58f51d679c742b8e5dde6aa6affd8745ab817fcd09477d129a81e74e07be576b5d3585eaca1c735b8e01 SHA512 804d25944035f33de6591fd942fbda44d3de037717a4397d38a97474b01775d30eaf93d16dd708a832c0119050d24d73b90990fd3e3773be79d26ada25244d22
 DIST pytorch-2.4.1.tar.gz 115029469 BLAKE2B c2909ff27d527bc57cba56b780d3b8cd07a043ab045caa6c6b27857a16f9ad10aaab2116b26226b1e46ee08ffb44007965d914464418e4ae14ca48c3f3f383bb SHA512 7e9b4485e242eaf0d648765c6621d73d95e7107b766646a098175436d1ab2e2b864badd0757a3bab6b7c318233f2120bad9ac07b39bb9e357897919580c87631
+DIST pytorch-2.5.1.tar.gz 116091366 BLAKE2B 7838b17562b94ffc7d798031348689db607dd5eae2a3c35be365972e2b52a2c1b12067068d5aca5ab00cf0977d9c2c3c9ae5337d69534c864c732e6256cbeef6 SHA512 a913a466324a65fa3d79c5e9ad4d605fc7976f0134fda2f81aaa3cea29d56926604999b8a238759646d211e63b47bbb446cdffa86ca8defd8159f11e30301289
diff --git a/sci-libs/caffe2/caffe2-2.5.1.ebuild b/sci-libs/caffe2/caffe2-2.5.1.ebuild
new file mode 100644
index 000000000000..81e3a916f3b3
--- /dev/null
+++ b/sci-libs/caffe2/caffe2-2.5.1.ebuild
@@ -0,0 +1,305 @@
+# Copyright 2022-2024 Gentoo Authors
+# Distributed under the terms of the GNU General Public License v2
+
+EAPI=8
+
+PYTHON_COMPAT=( python3_{10..12} )
+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 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-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-libs/onnx
+	sci-libs/foxi
+	cuda? (
+		dev-libs/cudnn
+		>=dev-libs/cudnn-frontend-1.0.3:0/8
+		<dev-util/nvidia-cuda-toolkit-12.5:=[profiler]
+	)
+	fbgemm? ( >=dev-libs/FBGEMM-2023.12.01 )
+	gloo? ( sci-libs/gloo[cuda?] )
+	mpi? ( virtual/mpi )
+	nnpack? ( sci-libs/NNPACK )
+	numpy? ( $(python_gen_cond_dep '
+		dev-python/numpy[${PYTHON_USEDEP}]
+		') )
+	onednn? ( dev-libs/oneDNN )
+	opencl? ( virtual/opencl )
+	qnnpack? (
+		!sci-libs/QNNPACK
+		dev-cpp/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-libs/tensorpipe[cuda?]
+		dev-cpp/cpp-httplib
+	)
+	xnnpack? ( >=sci-libs/XNNPACK-2024.02.29 )
+	mkl? ( sci-libs/mkl )
+	openblas? ( sci-libs/openblas )
+"
+DEPEND="
+	${RDEPEND}
+	cuda? ( >=dev-libs/cutlass-3.4.1 )
+	onednn? ( sci-libs/ideep )
+	dev-libs/psimd
+	dev-libs/FP16
+	dev-libs/FXdiv
+	dev-libs/pocketfft
+	dev-libs/flatbuffers
+	>=sci-libs/kineto-0.4.0_p20240525
+	$(python_gen_cond_dep '
+		dev-python/pyyaml[${PYTHON_USEDEP}]
+		dev-python/pybind11[${PYTHON_USEDEP}]
+		dev-python/typing-extensions[${PYTHON_USEDEP}]
+	')
+"
+
+PATCHES=(
+	"${FILESDIR}"/${P}-unbundle_fmt.patch
+	"${FILESDIR}"/${P}-unbundle_kineto.patch
+	"${FILESDIR}"/${P}-fix-functorch-install.patch
+	"${FILESDIR}"/${P}-cudnn_include_fix.patch
+	"${FILESDIR}"/${P}-gentoo.patch
+	"${FILESDIR}"/${P}-cpp-httplib.patch
+	"${FILESDIR}"/${P}-glog-0.6.0.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
+	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 ${TORCH_CUDA_ARCH_LIST} ]]; then
+		ewarn "WARNING: caffe2 is being built with its default CUDA compute capabilities: 3.5 and 7.0."
+		ewarn "These may not be optimal for your GPU."
+		ewarn ""
+		ewarn "To configure caffe2 with the CUDA compute capability that is optimal for your GPU,"
+		ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and re-emerge caffe2."
+		ewarn "For example, to use CUDA capability 7.5 & 3.5, add: TORCH_CUDA_ARCH_LIST=7.5 3.5"
+		ewarn "For a Maxwell model GPU, an example value would be: TORCH_CUDA_ARCH_LIST=Maxwell"
+		ewarn ""
+		ewarn "You can look up your GPU's CUDA compute capability at https://developer.nvidia.com/cuda-gpus"
+		ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery | grep 'CUDA Capability'"
+	fi
+
+	local mycmakeargs=(
+		-DLIBSHM_INSTALL_LIB_SUBDIR="${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=$(usex gloo)
+		-DUSE_ITT=OFF
+		-DUSE_KINETO=OFF # TODO
+		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
+		-DUSE_MEM_EFF_ATTENTION=OFF
+		-DUSE_MKLDNN=$(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=$(usex rocm)
+		-DUSE_SYSTEM_LIBS=ON
+		-DUSE_TENSORPIPE=$(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
+}
+
+src_install() {
+	cmake_src_install
+
+	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_domodule python/torch
+
+	dodir $(python_get_sitedir)/torch/bin
+	dodir $(python_get_sitedir)/torch/lib
+	dodir $(python_get_sitedir)/torch/include
+
+	ln -s ../../../../../include/torch \
+		"${D}$(python_get_sitedir)"/torch/include/torch || die # bug 923269
+
+
+	mv "${ED}"/usr/bin/torch_shm_manager \
+		"${ED}"/$(python_get_sitedir)/torch/bin/ || die
+
+	mv "${ED}"/usr/$(get_libdir)/libtorch_global_deps.so \
+		"${ED}"/$(python_get_sitedir)/torch/lib/ || die
+
+	mv "${ED}"/usr/lib/libc10*.so \
+		"${ED}"/usr/$(get_libdir)/ || die
+}
diff --git a/sci-libs/caffe2/files/caffe2-2.5.1-cpp-httplib.patch b/sci-libs/caffe2/files/caffe2-2.5.1-cpp-httplib.patch
new file mode 100644
index 000000000000..5d684a4a4738
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.5.1-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)
diff --git a/sci-libs/caffe2/files/caffe2-2.5.1-cudnn_include_fix.patch b/sci-libs/caffe2/files/caffe2-2.5.1-cudnn_include_fix.patch
new file mode 100644
index 000000000000..77905dbd1ac8
--- /dev/null
+++ b/sci-libs/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()
+ 
diff --git a/sci-libs/caffe2/files/caffe2-2.5.1-exclude-aotriton.patch b/sci-libs/caffe2/files/caffe2-2.5.1-exclude-aotriton.patch
new file mode 100644
index 000000000000..340d6e45c95a
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.5.1-exclude-aotriton.patch
@@ -0,0 +1,22 @@
+Disables aotriton download when both USE_FLASH_ATTENTION and USE_MEM_EFF_ATTENTION cmake flags are OFF
+Backports upstream PR to 2.3.0: https://github.com/pytorch/pytorch/pull/130197
+--- a/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp
++++ b/aten/src/ATen/native/transformers/cuda/sdp_utils.cpp
+@@ -659,7 +659,7 @@ bool can_use_mem_efficient_attention(sdp_params const& params, bool debug) {
+       array_of<at::ScalarType>(at::kHalf, at::kFloat, at::kBFloat16);
+   constexpr auto less_than_sm80_mem_efficient_dtypes =
+       array_of<at::ScalarType>(at::kHalf, at::kFloat);
+-#ifdef USE_ROCM
++#if defined(USE_ROCM) && defined(USE_MEM_EFF_ATTENTION)
+   constexpr auto aotriton_mem_efficient_dtypes =
+       array_of<at::ScalarType>(at::kHalf, at::kFloat, at::kBFloat16);
+ #endif
+@@ -709,7 +709,7 @@ bool can_use_mem_efficient_attention(sdp_params const& params, bool debug) {
+     }
+   }
+ 
+-#ifdef USE_ROCM
++#if defined(USE_ROCM) && defined(USE_MEM_EFF_ATTENTION)
+   return check_tensor_dtype(params, aotriton_mem_efficient_dtypes, debug);
+ #else
+   auto dprop = at::cuda::getCurrentDeviceProperties();
diff --git a/sci-libs/caffe2/files/caffe2-2.5.1-fix-functorch-install.patch b/sci-libs/caffe2/files/caffe2-2.5.1-fix-functorch-install.patch
new file mode 100644
index 000000000000..ffce0028a556
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.5.1-fix-functorch-install.patch
@@ -0,0 +1,8 @@
+--- a/functorch/CMakeLists.txt	2024-11-10 11:18:29.151992840 +0100
++++ b/functorch/CMakeLists.txt	2024-11-10 11:20:19.642389982 +0100
+@@ -42,4 +42,4 @@
+ if(NOT ${TORCH_PYTHON_LINK_FLAGS} STREQUAL "")
+   set_target_properties(${PROJECT_NAME} PROPERTIES LINK_FLAGS ${TORCH_PYTHON_LINK_FLAGS})
+ endif()
+-install(TARGETS ${PROJECT_NAME} DESTINATION "${CMAKE_CURRENT_SOURCE_DIR}")
++install(TARGETS ${PROJECT_NAME} DESTINATION "${CMAKE_INSTALL_LIBDIR}")
diff --git a/sci-libs/caffe2/files/caffe2-2.5.1-gentoo.patch b/sci-libs/caffe2/files/caffe2-2.5.1-gentoo.patch
new file mode 100644
index 000000000000..f923b6746a4b
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.5.1-gentoo.patch
@@ -0,0 +1,127 @@
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -986,12 +986,11 @@ endif()
+ # third_party/FBGEMM
+ include(cmake/public/utils.cmake)
+ if(NOT MSVC)
+-  string(APPEND CMAKE_CXX_FLAGS " -O2 -fPIC")
++  string(APPEND CMAKE_CXX_FLAGS " -O2")
+   # Eigen fails to build with some versions, so convert this to a warning
+   # 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)
+@@ -1085,7 +1084,6 @@ if(NOT MSVC)
+   string(APPEND CMAKE_LINKER_FLAGS_DEBUG " -fno-omit-frame-pointer -O0")
+   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)
+ else()
+   # skip unwanted includes from windows.h
+   add_compile_definitions(WIN32_LEAN_AND_MEAN)
+--- a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
++++ b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+@@ -324,16 +324,8 @@ set_target_properties(pytorch_qnnpack PROPERTIES PUBLIC_HEADER include/pytorch_q
+ set_target_properties(pytorch_qnnpack PROPERTIES PUBLIC_HEADER include/qnnpack_func.h)
+ 
+ # ---[ Configure clog
+-if(NOT TARGET clog)
+-  set(CLOG_BUILD_TESTS OFF CACHE BOOL "")
+-  set(CLOG_RUNTIME_TYPE "${CPUINFO_RUNTIME_TYPE}" CACHE STRING "")
+-  add_subdirectory(
+-    "${CLOG_SOURCE_DIR}"
+-    "${CONFU_DEPENDENCIES_BINARY_DIR}/clog")
+-  # We build static version of clog but a dynamic library may indirectly depend on it
+-  set_property(TARGET clog PROPERTY POSITION_INDEPENDENT_CODE ON)
+-endif()
+-target_link_libraries(pytorch_qnnpack PUBLIC clog)
++find_library(CLOG_LIBRARY NAMES clog REQUIRED)
++target_link_libraries(pytorch_qnnpack PUBLIC ${CLOG_LIBRARY})
+ 
+ # ---[ Configure cpuinfo
+ if(NOT TARGET cpuinfo AND USE_SYSTEM_CPUINFO)
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -87,7 +87,7 @@ endif()
+ # Note: the folders that are being commented out have not been properly
+ # addressed yet.
+ 
+-if(NOT MSVC AND USE_XNNPACK)
++if(FALSE)
+   if(NOT TARGET fxdiv)
+     set(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
+     set(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
+@@ -1081,7 +1081,6 @@ if(USE_XPU)
+ endif()
+ 
+ if(NOT MSVC AND USE_XNNPACK)
+-  TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)
+ endif()
+ 
+ # ==========================================================
+--- a/cmake/Codegen.cmake
++++ b/cmake/Codegen.cmake
+@@ -57,7 +57,7 @@ if(INTERN_BUILD_ATEN_OPS)
+   if(MSVC)
+     set(OPT_FLAG "/fp:strict ")
+   else(MSVC)
+-    set(OPT_FLAG "-O3 ")
++    set(OPT_FLAG " ")
+     if("${CMAKE_BUILD_TYPE}" MATCHES "Debug")
+       set(OPT_FLAG " ")
+     endif()
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -473,7 +473,9 @@ if(USE_PYTORCH_QNNPACK)
+       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 ${GEMMLOWP_INCLUDE_DIRS})
+ 
+       if(PYTORCH_QNNPACK_CUSTOM_THREADPOOL)
+         target_compile_definitions(
+@@ -710,7 +712,7 @@ if(BUILD_TEST OR BUILD_MOBILE_BENCHMARK OR BUILD_MOBILE_TEST)
+ endif()
+ 
+ # ---[ FBGEMM
+-if(USE_FBGEMM)
++if(FALSE)
+   set(CAFFE2_THIRD_PARTY_ROOT "${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")
+@@ -758,6 +760,7 @@ 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)
+--- a/cmake/External/nnpack.cmake
++++ b/cmake/External/nnpack.cmake
+@@ -56,7 +56,7 @@ if(ANDROID OR IOS OR ${CMAKE_SYSTEM_NAME} STREQUAL "Linux" OR ${CMAKE_SYSTEM_NAM
+   set(PTHREADPOOL_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/pthreadpool" CACHE STRING "pthreadpool source directory")
+   set(GOOGLETEST_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/googletest" CACHE STRING "Google Test source directory")
+ 
+-  if(NOT TARGET nnpack)
++  if(FALSE)
+     if(NOT USE_SYSTEM_PTHREADPOOL AND USE_INTERNAL_PTHREADPOOL_IMPL)
+       set(NNPACK_CUSTOM_THREADPOOL ON CACHE BOOL "")
+     endif()
+--- a/cmake/public/utils.cmake
++++ b/cmake/public/utils.cmake
+@@ -422,8 +422,6 @@ function(torch_compile_options libname)
+   endif()
+ 
+   # Use -O2 for release builds (-O3 doesn't improve perf, and -Os results in perf regression)
+-  target_compile_options(${libname} PRIVATE
+-      $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<OR:$<CONFIG:Release>,$<CONFIG:RelWithDebInfo>>>:-O2>)
+ 
+ endfunction()
+ 
diff --git a/sci-libs/caffe2/files/caffe2-2.5.1-glog-0.6.0.patch b/sci-libs/caffe2/files/caffe2-2.5.1-glog-0.6.0.patch
new file mode 100644
index 000000000000..6c06d2cca654
--- /dev/null
+++ b/sci-libs/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);
diff --git a/sci-libs/caffe2/files/caffe2-2.5.1-install-dirs.patch b/sci-libs/caffe2/files/caffe2-2.5.1-install-dirs.patch
new file mode 100644
index 000000000000..e99b7e59cb5b
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.5.1-install-dirs.patch
@@ -0,0 +1,11 @@
+--- a/c10/CMakeLists.txt
++++ b/c10/CMakeLists.txt
+@@ -162,7 +162,7 @@ if(NOT BUILD_LIBTORCHLESS)
+   # Note: for now, we will put all export path into one single Caffe2Targets group
+   # to deal with the cmake deployment need. Inside the Caffe2Targets set, the
+   # individual libraries like libc10.so and libcaffe2.so are still self-contained.
+-  install(TARGETS c10 EXPORT Caffe2Targets DESTINATION lib)
++  install(TARGETS c10 EXPORT Caffe2Targets DESTINATION ${CMAKE_INSTALL_LIBDIR})
+ endif()
+ 
+ install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR}
diff --git a/sci-libs/caffe2/files/caffe2-2.5.1-unbundle_fmt.patch b/sci-libs/caffe2/files/caffe2-2.5.1-unbundle_fmt.patch
new file mode 100644
index 000000000000..2594a56b36a5
--- /dev/null
+++ b/sci-libs/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)
diff --git a/sci-libs/caffe2/files/caffe2-2.5.1-unbundle_kineto.patch b/sci-libs/caffe2/files/caffe2-2.5.1-unbundle_kineto.patch
new file mode 100644
index 000000000000..ebe931bc49b6
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.5.1-unbundle_kineto.patch
@@ -0,0 +1,22 @@
+--- 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