sci-libs/caffe2: add 2.6.0

Signed-off-by: Alfredo Tupone <tupone@gentoo.org>
2026-09-24 04:59:14 -07:00 · 2025-02-27 21:53:15 +01:00
parent c6d531f550
commit 6435c53417
3 changed files with 483 additions and 0 deletions
--- 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-2.4.1.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
--- a/sci-libs/caffe2/caffe2-2.6.0.ebuild
+++ b/sci-libs/caffe2/caffe2-2.6.0.ebuild
@@ -0,0 +1,326 @@
+# 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 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-libs/onnx
+	sci-libs/foxi
+	cuda? (
+		dev-libs/cudnn
+		>=dev-libs/cudnn-frontend-1.0.3:0/8
+		dev-util/nvidia-cuda-toolkit:=[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-3.5* )
+	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}
+	qnnpack? ( dev-libs/clog )
+	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=(
+	"${FILESDIR}"/${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 $(get_libdir))|" \
+		c10/cuda/CMakeLists.txt \
+		c10/CMakeLists.txt \
+		c10/hip/CMakeLists.txt \
+		|| die
+
+	cmake_src_prepare
+	pushd torch/csrc/jit/serialization || die
+	flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
+	popd
+
+	# prefixify the hardcoded paths, after all patches are applied
+	hprefixify \
+		aten/CMakeLists.txt \
+		caffe2/CMakeLists.txt \
+		cmake/Metal.cmake \
+		cmake/Modules/*.cmake \
+		cmake/Modules_CUDA_fix/FindCUDNN.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
+		cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
+		cmake/public/LoadHIP.cmake \
+		cmake/public/cuda.cmake \
+		cmake/Dependencies.cmake \
+		torch/CMakeLists.txt \
+		CMakeLists.txt
+
+	if use rocm; then
+		sed -e "s:/opt/rocm:/usr:" \
+			-e "s:lib/cmake:$(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=$(usex gloo)
+		-DUSE_ITT=OFF
+		-DUSE_KINETO=OFF # TODO
+		-DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
+		-DUSE_MEM_EFF_ATTENTION=OFF
+		-DUSE_MKLDNN=$(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=$(usex rocm)
+		-DUSE_SYSTEM_CPUINFO=ON
+		-DUSE_SYSTEM_EIGEN_INSTALL=ON
+		-DUSE_SYSTEM_FP16=ON
+		-DUSE_SYSTEM_FXDIV=ON
+		-DUSE_SYSTEM_GLOO=ON
+		-DUSE_SYSTEM_ONNX=ON
+		-DUSE_SYSTEM_PSIMD=ON
+		-DUSE_SYSTEM_PSIMD=ON
+		-DUSE_SYSTEM_PTHREADPOOL=ON
+		-DUSE_SYSTEM_PYBIND11=ON
+		-DUSE_SYSTEM_SLEEF=ON
+		-DUSE_SYSTEM_XNNPACK=$(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=${PV} \
+	PYTORCH_BUILD_NUMBER=0 \
+	cmake_src_compile
+}
+
+src_install() {
+	cmake_src_install
+
+	# Used by pytorch ebuild
+	insinto "/var/lib/${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_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
+
+	mv "${D}"/usr/bin/torch_shm_manager \
+		"${D}"/$(python_get_sitedir)/torch/bin/ || die
+
+	mv "${D}"/usr/$(get_libdir)/libtorch_global_deps.so \
+		"${D}"/$(python_get_sitedir)/torch/lib/ || die
+}
--- a/sci-libs/caffe2/files/caffe2-2.6.0-gentoo.patch
+++ b/sci-libs/caffe2/files/caffe2-2.6.0-gentoo.patch
@@ -0,0 +1,156 @@
+--- a/CMakeLists.txt
+++ b/CMakeLists.txt
+@@ -984,12 +984,11 @@ endif()
+ # third_party/FBGEMM
+ include(cmake/public/utils.cmake)
+ if(NOT MSVC)
+-  string(APPEND CMAKE_CXX_FLAGS " -O2 -fPIC")
+  string(APPEND CMAKE_CXX_FLAGS " -O2")
+   # Eigen fails to build with some versions, so convert this to a warning
+   # Details at http://eigen.tuxfamily.org/bz/show_bug.cgi?id=1459
+   string(APPEND CMAKE_CXX_FLAGS " -Wall")
+   string(APPEND CMAKE_CXX_FLAGS " -Wextra")
+-  append_cxx_flag_if_supported("-Werror=return-type" CMAKE_CXX_FLAGS)
+   append_cxx_flag_if_supported("-Werror=non-virtual-dtor" CMAKE_CXX_FLAGS)
+   append_cxx_flag_if_supported("-Werror=braced-scalar-init" CMAKE_CXX_FLAGS)
+   append_cxx_flag_if_supported("-Werror=range-loop-construct" CMAKE_CXX_FLAGS)
+@@ -1091,7 +1091,6 @@
+   endif()
+   append_cxx_flag_if_supported("-fno-math-errno" CMAKE_CXX_FLAGS)
+   append_cxx_flag_if_supported("-fno-trapping-math" CMAKE_CXX_FLAGS)
+-  append_cxx_flag_if_supported("-Werror=format" CMAKE_CXX_FLAGS)
+   if(CMAKE_COMPILER_IS_GNUCXX AND CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 13)
+     append_cxx_flag_if_supported("-Wno-error=dangling-reference" CMAKE_CXX_FLAGS)
+     append_cxx_flag_if_supported("-Wno-error=redundant-move" CMAKE_CXX_FLAGS)
+--- a/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+++ b/aten/src/ATen/native/quantized/cpu/qnnpack/CMakeLists.txt
+@@ -323,16 +323,8 @@ set_target_properties(pytorch_qnnpack PROPERTIES PUBLIC_HEADER include/pytorch_q
+ set_target_properties(pytorch_qnnpack PROPERTIES PUBLIC_HEADER include/qnnpack_func.h)
+ 
+ # ---[ Configure clog
+-if(NOT TARGET clog)
+-  set(CLOG_BUILD_TESTS OFF CACHE BOOL "")
+-  set(CLOG_RUNTIME_TYPE "${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 ${CLOG_LIBRARY})
+ 
+ # ---[ Configure cpuinfo
+ if(NOT TARGET cpuinfo AND USE_SYSTEM_CPUINFO)
+--- a/caffe2/CMakeLists.txt
+++ b/caffe2/CMakeLists.txt
+@@ -87,7 +87,7 @@ endif()
+ # Note: the folders that are being commented out have not been properly
+ # addressed yet.
+ 
+-if(NOT MSVC AND USE_XNNPACK)
+if(FALSE)
+   if(NOT TARGET fxdiv)
+     set(FXDIV_BUILD_TESTS OFF CACHE BOOL "")
+     set(FXDIV_BUILD_BENCHMARKS OFF CACHE BOOL "")
+@@ -1130,7 +1130,6 @@ if(USE_XPU)
+ endif()
+ 
+ if(NOT MSVC AND USE_XNNPACK)
+-  TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)
+ endif()
+ 
+ # ==========================================================
+--- a/cmake/Codegen.cmake
+++ b/cmake/Codegen.cmake
+@@ -64,7 +64,7 @@ if(INTERN_BUILD_ATEN_OPS)
+   if(MSVC)
+     set(OPT_FLAG "/fp:strict ")
+   else(MSVC)
+-    set(OPT_FLAG "-O3 ")
+    set(OPT_FLAG " ")
+     if("${CMAKE_BUILD_TYPE}" MATCHES "Debug")
+       set(OPT_FLAG " ")
+     endif()
+--- a/cmake/Dependencies.cmake
+++ b/cmake/Dependencies.cmake
+@@ -467,7 +467,9 @@
+       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 ${GEMMLOWP_INCLUDE_DIRS})
+     endif()
+ 
+     list(APPEND Caffe2_DEPENDENCY_LIBS pytorch_qnnpack)
+@@ -556,7 +556,7 @@
+   find_library(microkernels-prod_LIBRARY microkernels-prod)
+   set_property(TARGET XNNPACK PROPERTY IMPORTED_LOCATION "${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: ${XNNPACK_LIBRARY}")
+@@ -693,7 +695,7 @@ if(BUILD_TEST OR BUILD_MOBILE_BENCHMARK OR BUILD_MOBILE_TEST)
+ endif()
+ 
+ # ---[ FBGEMM
+-if(USE_FBGEMM)
+if(FALSE)
+   set(CAFFE2_THIRD_PARTY_ROOT "${PROJECT_SOURCE_DIR}/third_party")
+   if(NOT DEFINED FBGEMM_SOURCE_DIR)
+     set(FBGEMM_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/fbgemm" CACHE STRING "FBGEMM source directory")
+@@ -745,6 +747,7 @@ if(USE_FBGEMM)
+ endif()
+ 
+ if(USE_FBGEMM)
+  list(APPEND Caffe2_DEPENDENCY_LIBS fbgemm)
+   caffe2_update_option(USE_FBGEMM ON)
+ else()
+   caffe2_update_option(USE_FBGEMM OFF)
+--- a/cmake/External/nnpack.cmake
+++ b/cmake/External/nnpack.cmake
+@@ -56,7 +56,7 @@
+   set(PTHREADPOOL_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/pthreadpool" CACHE STRING "pthreadpool source directory")
+   set(GOOGLETEST_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/googletest" CACHE STRING "Google Test source directory")
+ 
+-  if(NOT TARGET nnpack)
+  if(FALSE)
+     set(NNPACK_BUILD_TESTS OFF CACHE BOOL "")
+     set(NNPACK_BUILD_BENCHMARKS OFF CACHE BOOL "")
+     set(NNPACK_LIBRARY_TYPE "static" CACHE STRING "")
+--- a/cmake/public/utils.cmake
+++ b/cmake/public/utils.cmake
+@@ -422,8 +422,6 @@ function(torch_compile_options libname)
+   endif()
+ 
+   # Use -O2 for release builds (-O3 doesn't improve perf, and -Os results in perf regression)
+-  target_compile_options(${libname} PRIVATE
+-      $<$<AND:$<COMPILE_LANGUAGE:CXX>,$<OR:$<CONFIG:Release>,$<CONFIG:RelWithDebInfo>>>:-O2>)
+ 
+ endfunction()
+ 
+--- 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
+@@ -269,8 +269,6 @@
+ 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}
+@@ -315,8 +313,6 @@
+   list(APPEND ATen_HIP_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/composable_kernel/include)
+   list(APPEND ATen_HIP_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/composable_kernel/library/include)
+   # 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)
+   list(APPEND ATen_HIP_SRCS
+     ${ATen_HIP_SRCS}
+     ${hip_hip}