commit: 7354afd9d872a15c6bdf1ee831db7f7f131fca01
Author: Sv. Lockal <lockalsash <AT> gmail <DOT> com>
AuthorDate: Thu Aug 8 03:44:50 2024 +0000
Commit: Alfredo Tupone <tupone <AT> gentoo <DOT> org>
CommitDate: Thu Aug 8 16:22:28 2024 +0000
URL: https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=7354afd9
sci-libs/caffe2: add 2.4.0
Changes:
* This version has no integration with ffmpeg/opencv anymore
* qnnpack ceased to exist as a separate library. Now it is a part of c10
sources. Splitting it to sci-libs/QNNPACK would require some major effort,
while it is very unlikely that some non-pytorch project would use it. Therefore
it is now bundled and sci-libs/QNNPACK added as a conflicting library (to be
sunset in the future).
* sci-libs/hipBLASLt "is optional" patch was dropped, as it is now deeply
integrated with caffe2. However it should not be a problem as
sci-libs/hipBLASLt has "no kernels" mode where it compiles as a stub library
(useful for all non-server GPUs)
* sci-libs/XNNPACK bumped; old xnnpack is not compatible
* sci-libs/kineto bumped; old kineto is not compatible
* no new patches, but 5 old patches require manual conflict resolution
* '-fclang-abi-compat=17' patch in previous release was incomplete; new version
should handle clang+hipcc combination
Signed-off-by: Sv. Lockal <lockalsash <AT> gmail.com>
Signed-off-by: Alfredo Tupone <tupone <AT> gentoo.org>
sci-libs/caffe2/Manifest | 1 +
sci-libs/caffe2/caffe2-2.4.0.ebuild | 289 +++++++++++++++++++++
.../files/caffe2-2.4.0-exclude-aotriton.patch | 65 +++++
.../files/caffe2-2.4.0-fix-openmp-link.patch | 14 +
sci-libs/caffe2/files/caffe2-2.4.0-gentoo.patch | 211 +++++++++++++++
.../caffe2/files/caffe2-2.4.0-install-dirs.patch | 70 +++++
.../files/caffe2-2.4.0-rocm-fix-std-cpp17.patch | 50 ++++
7 files changed, 700 insertions(+)
diff --git a/sci-libs/caffe2/Manifest b/sci-libs/caffe2/Manifest
index 246ab4d21a1a..f34355be2e16 100644
--- a/sci-libs/caffe2/Manifest
+++ b/sci-libs/caffe2/Manifest
@@ -1,3 +1,4 @@
DIST pytorch-2.2.2.tar.gz 116367503 BLAKE2B
0be22f2ec4b9aac6f5e976664cae01facf07929a32565cd57d7cc5b2d9888e9ae71ca301853752fe8f31d174d04c9974eb9ed2f3d452360a50ccf024f200726a
SHA512
7990e0f9484038c3458c0bda2c863bf2b19e56edab81fc5938c6e0f08b17558287f853bb67350e8cca8f42bec0f1d4ba0e94e50a145db8da44bdd4bd703d91d0
DIST pytorch-2.3.0.tar.gz 117029829 BLAKE2B
8f9c0d71ee0a9219b495eddccdcc65107f7ad537c43c68100b229f3d27b0e6c01ccb1659c7fffc356a48d80f2adc0a10361305dc8f1df20446de837d380f89f6
SHA512
67f7e9a096c3ffb952206ebf9105bedebb68c24ad82456083adf1d1d210437fcaa9dd52b68484cfc97d408c9eebc9541c76868c34a7c9982494dc3f424cfb07c
DIST pytorch-2.3.1.tar.gz 117035696 BLAKE2B
d419d7fa1342f1fb317ffce09ec9dc1447414627cc83d36578fe60f68c283c620b2b4d49f414cd206d537b90b16432a06cd1941662720db05d5e2b6c493325f5
SHA512
e1bcae44f9939fc7ccb1360a9b1970d92426f25e5de73e36964df3dd15ad5d8d9f5bd2f9a7dda6b8f64e2bba3674005bd869f542489cc442ad0125a02676f587
+DIST pytorch-2.4.0.tar.gz 115031093 BLAKE2B
d206477963977011627df284efa01482fbf57e9fcb5f58f51d679c742b8e5dde6aa6affd8745ab817fcd09477d129a81e74e07be576b5d3585eaca1c735b8e01
SHA512
804d25944035f33de6591fd942fbda44d3de037717a4397d38a97474b01775d30eaf93d16dd708a832c0119050d24d73b90990fd3e3773be79d26ada25244d22
diff --git a/sci-libs/caffe2/caffe2-2.4.0.ebuild
b/sci-libs/caffe2/caffe2-2.4.0.ebuild
new file mode 100644
index 000000000000..b4384eb7df11
--- /dev/null
+++ b/sci-libs/caffe2/caffe2-2.4.0.ebuild
@@ -0,0 +1,289 @@
+# 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
+ )
+"
+
+# CUDA 12 not supported yet: https://github.com/pytorch/pytorch/issues/91122
+RDEPEND="
+ ${PYTHON_DEPS}
+ 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
+ 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-12.4.0:=[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?] )
+ 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=(
+ "${FILESDIR}"/${PN}-2.4.0-gentoo.patch
+ "${FILESDIR}"/${PN}-2.4.0-install-dirs.patch
+ "${FILESDIR}"/${PN}-1.12.0-glog-0.6.0.patch
+ "${FILESDIR}"/${PN}-1.13.1-tensorpipe.patch
+ "${FILESDIR}"/${PN}-2.3.0-cudnn_include_fix.patch
+ "${FILESDIR}"/${PN}-2.1.2-fix-rpath.patch
+ "${FILESDIR}"/${PN}-2.4.0-fix-openmp-link.patch
+ "${FILESDIR}"/${PN}-2.4.0-rocm-fix-std-cpp17.patch
+ "${FILESDIR}"/${PN}-2.2.2-musl.patch
+ "${FILESDIR}"/${PN}-2.4.0-exclude-aotriton.patch
+ "${FILESDIR}"/${PN}-2.3.0-fix-rocm-gcc14-clamp.patch
+ "${FILESDIR}"/${PN}-2.3.0-fix-libcpp.patch
+)
+
+src_prepare() {
+ filter-lto #bug 862672
+ sed -i \
+ -e "/third_party\/gloo/d" \
+ cmake/Dependencies.cmake \
+ || die
+ cmake_src_prepare
+ pushd torch/csrc/jit/serialization || die
+ flatc --cpp --gen-mutable --scoped-enums mobile_bytecode.fbs || die
+ popd
+ # prefixify the hardcoded paths, after all patches are applied
+ hprefixify \
+ aten/CMakeLists.txt \
+ caffe2/CMakeLists.txt \
+ cmake/Metal.cmake \
+ cmake/Modules/*.cmake \
+ cmake/Modules_CUDA_fix/FindCUDNN.cmake \
+ cmake/Modules_CUDA_fix/upstream/FindCUDA/make2cmake.cmake \
+
cmake/Modules_CUDA_fix/upstream/FindPackageHandleStandardArgs.cmake \
+ cmake/public/LoadHIP.cmake \
+ cmake/public/cuda.cmake \
+ cmake/Dependencies.cmake \
+ torch/CMakeLists.txt \
+ CMakeLists.txt
+
+ if use rocm; then
+ sed -e "s:/opt/rocm:/usr:" \
+ -e "s:lib/cmake:$(get_libdir)/cmake:g" \
+ -e "s/HIP 1.0/HIP 1.0 REQUIRED/" \
+ -i cmake/public/LoadHIP.cmake || die
+
+ ebegin "HIPifying cuda sources"
+ ${EPYTHON} tools/amd_build/build_amd.py || die
+ eend $?
+ fi
+}
+
+src_configure() {
+ if use cuda && [[ -z ${TORCH_CUDA_ARCH_LIST} ]]; then
+ ewarn "WARNING: caffe2 is being built with its default CUDA
compute capabilities: 3.5 and 7.0."
+ ewarn "These may not be optimal for your GPU."
+ ewarn ""
+ ewarn "To configure caffe2 with the CUDA compute capability
that is optimal for your GPU,"
+ ewarn "set TORCH_CUDA_ARCH_LIST in your make.conf, and
re-emerge caffe2."
+ ewarn "For example, to use CUDA capability 7.5 & 3.5, add:
TORCH_CUDA_ARCH_LIST=7.5 3.5"
+ ewarn "For a Maxwell model GPU, an example value would be:
TORCH_CUDA_ARCH_LIST=Maxwell"
+ ewarn ""
+ ewarn "You can look up your GPU's CUDA compute capability at
https://developer.nvidia.com/cuda-gpus";
+ ewarn "or by running /opt/cuda/extras/demo_suite/deviceQuery |
grep 'CUDA Capability'"
+ fi
+
+ local mycmakeargs=(
+ -DBUILD_CUSTOM_PROTOBUF=OFF
+ -DBUILD_SHARED_LIBS=ON
+
+ -DUSE_CCACHE=OFF
+ -DUSE_CUDA=$(usex cuda)
+ -DUSE_DISTRIBUTED=$(usex distributed)
+ -DUSE_MPI=$(usex mpi)
+ -DUSE_FAKELOWP=OFF
+ -DUSE_FBGEMM=$(usex fbgemm)
+ -DUSE_FLASH_ATTENTION=$(usex flash)
+ -DUSE_MEM_EFF_ATTENTION=OFF
+ -DUSE_GFLAGS=ON
+ -DUSE_GLOG=ON
+ -DUSE_GLOO=$(usex gloo)
+ -DUSE_KINETO=OFF # TODO
+ -DUSE_MAGMA=OFF # TODO: In GURU as sci-libs/magma
+ -DUSE_MKLDNN=$(usex onednn)
+ -DUSE_NNPACK=$(usex nnpack)
+ -DUSE_XNNPACK=$(usex xnnpack)
+ -DUSE_SYSTEM_XNNPACK=$(usex xnnpack)
+ -DUSE_TENSORPIPE=$(usex distributed)
+ -DUSE_PYTORCH_QNNPACK=$(usex qnnpack)
+ -DUSE_NUMPY=$(usex numpy)
+ -DUSE_OPENCL=$(usex opencl)
+ -DUSE_OPENMP=$(usex openmp)
+ -DUSE_ROCM=$(usex rocm)
+ -DUSE_SYSTEM_CPUINFO=ON
+ -DUSE_SYSTEM_PYBIND11=ON
+ -DUSE_UCC=OFF
+ -DUSE_VALGRIND=OFF
+ -DPython_EXECUTABLE="${PYTHON}"
+ -DUSE_ITT=OFF
+ -DUSE_SYSTEM_PTHREADPOOL=ON
+ -DUSE_SYSTEM_PSIMD=ON
+ -DUSE_SYSTEM_FXDIV=ON
+ -DUSE_SYSTEM_FP16=ON
+ -DUSE_SYSTEM_GLOO=ON
+ -DUSE_SYSTEM_ONNX=ON
+ -DUSE_SYSTEM_SLEEF=ON
+ -DUSE_PYTORCH_METAL=OFF
+ -DUSE_XPU=OFF
+
+ -Wno-dev
+ -DTORCH_INSTALL_LIB_DIR="${EPREFIX}"/usr/$(get_libdir)
+ -DLIBSHM_INSTALL_LIB_SUBDIR="${EPREFIX}"/usr/$(get_libdir)
+ )
+
+ if use mkl; then
+ mycmakeargs+=(-DBLAS=MKL)
+ elif use openblas; then
+ mycmakeargs+=(-DBLAS=OpenBLAS)
+ else
+ mycmakeargs+=(-DBLAS=Generic -DBLAS_LIBRARIES=)
+ fi
+
+ if use cuda; then
+ addpredict "/dev/nvidiactl" # bug 867706
+ addpredict "/dev/char"
+ addpredict "/proc/self/task" # bug 926116
+
+ mycmakeargs+=(
+ -DUSE_CUDNN=ON
+ -DTORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST:-3.5
7.0}"
+ -DUSE_NCCL=OFF # TODO: NVIDIA Collective Communication
Library
+ -DCMAKE_CUDA_FLAGS="$(cuda_gccdir -f | tr -d \")"
+ )
+ elif use rocm; then
+ export PYTORCH_ROCM_ARCH="$(get_amdgpu_flags)"
+
+ mycmakeargs+=(
+ -DUSE_NCCL=ON
+ -DUSE_SYSTEM_NCCL=ON
+ )
+
+ # ROCm libraries produce too much warnings
+ append-cxxflags -Wno-deprecated-declarations -Wno-unused-result
+
+ if tc-is-clang; then
+ # fix mangling in LLVM:
https://github.com/llvm/llvm-project/issues/85656
+ append-cxxflags -fclang-abi-compat=17
+ fi
+ fi
+
+ if use onednn; then
+ mycmakeargs+=(
+ -DUSE_MKLDNN=ON
+ -DMKLDNN_FOUND=ON
+ -DMKLDNN_LIBRARIES=dnnl
+
-DMKLDNN_INCLUDE_DIR="${ESYSROOT}/usr/include/oneapi/dnnl"
+ )
+ fi
+
+ cmake_src_configure
+
+ # do not rerun cmake and the build process in src_install
+ sed '/RERUN/,+1d' -i "${BUILD_DIR}"/build.ninja || die
+}
+
+src_install() {
+ cmake_src_install
+
+ insinto "/var/lib/${PN}"
+ doins "${BUILD_DIR}"/CMakeCache.txt
+
+ rm -rf python
+ mkdir -p python/torch/include || die
+ mv "${ED}"/usr/lib/python*/site-packages/caffe2 python/ || die
+ cp torch/version.py python/torch/ || die
+ python_domodule python/caffe2
+ python_domodule python/torch
+ ln -s ../../../../../include/torch \
+ "${D}$(python_get_sitedir)"/torch/include/torch || die # bug
923269
+}
diff --git a/sci-libs/caffe2/files/caffe2-2.4.0-exclude-aotriton.patch
b/sci-libs/caffe2/files/caffe2-2.4.0-exclude-aotriton.patch
new file mode 100644
index 000000000000..72ab792b2278
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.4.0-exclude-aotriton.patch
@@ -0,0 +1,65 @@
+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
+@@ -24,7 +24,7 @@
+ #include <c10/core/SymInt.h>
+ #include <c10/util/string_view.h>
+
+-#if USE_ROCM
++#if defined(USE_ROCM) && (defined(USE_MEM_EFF_ATTENTION) ||
defined(USE_FLASH_ATTENTION))
+ #include <aotriton/flash.h>
+ #endif
+
+@@ -207,7 +207,7 @@ bool check_flash_attention_hardware_support(sdp_params
const& params, bool debug
+ // Check that the gpu is capable of running flash attention
+ using sm80 = SMVersion<8, 0>;
+ using sm90 = SMVersion<9, 0>;
+-#if USE_ROCM
++#if defined(USE_ROCM) && defined(USE_FLASH_ATTENTION)
+ auto stream = at::cuda::getCurrentCUDAStream().stream();
+ if (hipSuccess != aotriton::v2::flash::check_gpu(stream)) {
+ auto dprops = at::cuda::getCurrentDeviceProperties();
+@@ -238,7 +238,7 @@ bool check_mem_efficient_hardware_support(sdp_params
const& params, bool debug)
+ // Mem Efficient attention supports hardware in the range [sm_50, sm_90]
+ using sm50 = SMVersion<5, 0>;
+ using sm90 = SMVersion<9, 0>;
+-#if USE_ROCM
++#if defined(USE_ROCM) && defined(USE_MEM_EFF_ATTENTION)
+ auto stream = at::cuda::getCurrentCUDAStream().stream();
+ if (hipSuccess != aotriton::v2::flash::check_gpu(stream)) {
+ auto dprops = at::cuda::getCurrentDeviceProperties();
+@@ -623,7 +623,7 @@ 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
+@@ -668,7 +668,7 @@ 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();
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -1095,10 +1095,12 @@ if(USE_ROCM)
+ message(STATUS "Disabling Kernel Assert for ROCm")
+ endif()
+
+- include(${CMAKE_CURRENT_LIST_DIR}/External/aotriton.cmake)
+ if(USE_CUDA)
+ caffe2_update_option(USE_MEM_EFF_ATTENTION OFF)
+ endif()
++ if(USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION)
++ include(${CMAKE_CURRENT_LIST_DIR}/External/aotriton.cmake)
++ endif()
+ else()
+ caffe2_update_option(USE_ROCM OFF)
+ endif()
diff --git a/sci-libs/caffe2/files/caffe2-2.4.0-fix-openmp-link.patch
b/sci-libs/caffe2/files/caffe2-2.4.0-fix-openmp-link.patch
new file mode 100644
index 000000000000..9b0fe0b97c0f
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.4.0-fix-openmp-link.patch
@@ -0,0 +1,14 @@
+Fix "undefined symbol: omp_get_max_active_levels" in mkl + <nothing else>
builds
+https://github.com/pytorch/pytorch/issues/116576
+--- a/caffe2/CMakeLists.txt
++++ b/caffe2/CMakeLists.txt
+@@ -1643,6 +1643,9 @@ if(BUILD_SHARED_LIBS)
+ if(CAFFE2_USE_MKL)
+ target_link_libraries(torch_global_deps caffe2::mkl)
+ endif()
++ if(USE_OPENMP)
++ target_link_libraries(torch_global_deps OpenMP::OpenMP_CXX)
++ endif()
+ # The CUDA libraries are linked here for a different reason: in some
+ # cases we load these libraries with ctypes, and if they weren't opened
+ # with RTLD_GLOBAL, we'll do the "normal" search process again (and
diff --git a/sci-libs/caffe2/files/caffe2-2.4.0-gentoo.patch
b/sci-libs/caffe2/files/caffe2-2.4.0-gentoo.patch
new file mode 100644
index 000000000000..d592a346386b
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.4.0-gentoo.patch
@@ -0,0 +1,211 @@
+--- a/CMakeLists.txt
++++ b/CMakeLists.txt
+@@ -966,12 +966,11 @@ 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)
+@@ -1074,7 +1073,6 @@ 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/c10/CMakeLists.txt
++++ b/c10/CMakeLists.txt
+@@ -94,7 +94,7 @@ if(NOT BUILD_LIBTORCHLESS)
+ if(C10_USE_GLOG)
+ target_link_libraries(c10 PUBLIC glog::glog)
+ endif()
+- target_link_libraries(c10 PRIVATE fmt::fmt-header-only)
++ target_link_libraries(c10 PRIVATE fmt)
+
+ if(C10_USE_NUMA)
+ message(STATUS "NUMA paths:")
+--- 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 "")
+@@ -1075,7 +1075,6 @@ if(USE_XPU)
+ endif()
+
+ if(NOT MSVC AND USE_XNNPACK)
+- TARGET_LINK_LIBRARIES(torch_cpu PRIVATE fxdiv)
+ endif()
+
+ # ==========================================================
+@@ -1178,8 +1177,7 @@ target_include_directories(torch_cpu PRIVATE
+ target_include_directories(torch_cpu PRIVATE
+ ${TORCH_ROOT}/third_party/miniz-2.1.0)
+
+-target_include_directories(torch_cpu PRIVATE
+- ${TORCH_ROOT}/third_party/kineto/libkineto/include)
++target_include_directories(torch_cpu PRIVATE /usr/include/kineto)
+
+ if(USE_KINETO)
+ target_include_directories(torch_cpu PRIVATE
+--- a/cmake/Codegen.cmake
++++ b/cmake/Codegen.cmake
+@@ -57,7 +57,7 @@ if(INTERN_BUILD_ATEN_OPS)
+ if(MSVC)
+ set(OPT_FLAG "/fp:strict ")
+ else(MSVC)
+- set(OPT_FLAG "-O3 ")
++ set(OPT_FLAG " ")
+ if("${CMAKE_BUILD_TYPE}" MATCHES "Debug")
+ set(OPT_FLAG " ")
+ endif()
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -466,7 +466,9 @@ 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
${GEMMLOWP_INCLUDE_DIRS})
+
+ if(PYTORCH_QNNPACK_CUSTOM_THREADPOOL)
+ target_compile_definitions(
+@@ -705,7 +707,7 @@ if(BUILD_TEST OR BUILD_MOBILE_BENCHMARK OR
BUILD_MOBILE_TEST)
+ endif()
+
+ # ---[ FBGEMM
+-if(USE_FBGEMM)
++if(FALSE)
+ set(CAFFE2_THIRD_PARTY_ROOT "${PROJECT_SOURCE_DIR}/third_party")
+ if(NOT DEFINED FBGEMM_SOURCE_DIR)
+ set(FBGEMM_SOURCE_DIR "${CAFFE2_THIRD_PARTY_ROOT}/fbgemm" CACHE STRING
"FBGEMM source directory")
+@@ -753,6 +755,7 @@ if(USE_FBGEMM)
+ endif()
+
+ if(USE_FBGEMM)
++ list(APPEND Caffe2_DEPENDENCY_LIBS fbgemm)
+ caffe2_update_option(USE_FBGEMM ON)
+ else()
+ caffe2_update_option(USE_FBGEMM OFF)
+@@ -1288,7 +1291,6 @@ if(CAFFE2_CMAKE_BUILDING_WITH_MAIN_REPO AND NOT
INTERN_DISABLE_ONNX)
+ set_target_properties(onnx_proto PROPERTIES CXX_STANDARD 17)
+ endif()
+ endif()
+- add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/../third_party/foxi
EXCLUDE_FROM_ALL)
+
+ add_definitions(-DONNX_NAMESPACE=${ONNX_NAMESPACE})
+ if(NOT USE_SYSTEM_ONNX)
+@@ -1530,7 +1532,6 @@ endif()
+ #
+ set(TEMP_BUILD_SHARED_LIBS ${BUILD_SHARED_LIBS})
+ set(BUILD_SHARED_LIBS OFF CACHE BOOL "Build shared libs" FORCE)
+-add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)
+
+ # Disable compiler feature checks for `fmt`.
+ #
+@@ -1539,9 +1540,7 @@ add_subdirectory(${PROJECT_SOURCE_DIR}/third_party/fmt)
+ # 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::fmt-header-only)
+ set(BUILD_SHARED_LIBS ${TEMP_BUILD_SHARED_LIBS} CACHE BOOL "Build shared
libs" FORCE)
+
+ # ---[ Kineto
+--- 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 "${CAFFE2_THIRD_PARTY_ROOT}/googletest" CACHE
STRING "Google Test source directory")
+
+- if(NOT TARGET nnpack)
++ if(FALSE)
+ if(NOT USE_SYSTEM_PTHREADPOOL AND USE_INTERNAL_PTHREADPOOL_IMPL)
+ set(NNPACK_CUSTOM_THREADPOOL ON CACHE BOOL "")
+ endif()
+--- a/cmake/public/utils.cmake
++++ b/cmake/public/utils.cmake
+@@ -483,8 +483,6 @@ function(torch_compile_options libname)
+ endif()
+
+ # Use -O2 for release builds (-O3 doesn't improve perf, and -Os results in
perf regression)
+- target_compile_options(${libname} PRIVATE
+-
$<$<AND:$<COMPILE_LANGUAGE:CXX>,$<OR:$<CONFIG:Release>,$<CONFIG:RelWithDebInfo>>>:-O2>)
+
+ endfunction()
+
+--- a/functorch/CMakeLists.txt
++++ b/functorch/CMakeLists.txt
+@@ -42,4 +42,4 @@ endif()
+ 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}")
+--- a/torch/CMakeLists.txt
++++ b/torch/CMakeLists.txt
+@@ -59,16 +59,8 @@ set(TORCH_PYTHON_INCLUDE_DIRECTORIES
+ ${CMAKE_BINARY_DIR}
+ ${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
+- ${TORCH_ROOT}/third_party/kineto/libkineto/include
+- ${TORCH_ROOT}/third_party/cpp-httplib
++
++ /usr/include/kineto
+
+ ${TORCH_SRC_DIR}/csrc
+ ${TORCH_SRC_DIR}/csrc/api/include
+@@ -83,7 +75,6 @@ set(TORCH_PYTHON_LINK_LIBRARIES
+ opentelemetry::api
+ httplib
+ shm
+- fmt::fmt-header-only
+ ATEN_CPU_FILES_GEN_LIB)
+
+ if(USE_ASAN AND TARGET Sanitizer::address)
diff --git a/sci-libs/caffe2/files/caffe2-2.4.0-install-dirs.patch
b/sci-libs/caffe2/files/caffe2-2.4.0-install-dirs.patch
new file mode 100644
index 000000000000..ee6e8fb91562
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.4.0-install-dirs.patch
@@ -0,0 +1,70 @@
+--- a/c10/CMakeLists.txt
++++ b/c10/CMakeLists.txt
+@@ -157,7 +157,7 @@ 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}
+--- a/c10/cuda/CMakeLists.txt
++++ b/c10/cuda/CMakeLists.txt
+@@ -82,7 +82,7 @@ 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_cuda EXPORT Caffe2Targets DESTINATION lib)
++install(TARGETS c10_cuda EXPORT Caffe2Targets DESTINATION
${CMAKE_INSTALL_LIBDIR})
+
+ endif()
+
+--- a/c10/hip/CMakeLists.txt
++++ b/c10/hip/CMakeLists.txt
+@@ -57,7 +57,7 @@ if(NOT BUILD_LIBTORCHLESS)
+ $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}/../..>
+ $<BUILD_INTERFACE:${CMAKE_BINARY_DIR}>
+ $<INSTALL_INTERFACE:include>)
+- install(TARGETS c10_hip EXPORT Caffe2Targets DESTINATION lib)
++ install(TARGETS c10_hip EXPORT Caffe2Targets DESTINATION
${CMAKE_INSTALL_LIBDIR})
+ set(C10_HIP_LIB c10_hip)
+ endif()
+
+--- a/c10/xpu/CMakeLists.txt
++++ b/c10/xpu/CMakeLists.txt
+@@ -45,7 +45,7 @@ target_include_directories(
+ $<BUILD_INTERFACE:${CMAKE_BINARY_DIR}>
+ $<INSTALL_INTERFACE:include>
+ )
+- install(TARGETS c10_xpu EXPORT Caffe2Targets DESTINATION lib)
++ install(TARGETS c10_xpu EXPORT Caffe2Targets DESTINATION
${CMAKE_INSTALL_LIBDIR})
+ set(C10_XPU_LIB c10_xpu)
+ add_subdirectory(test)
+ endif()
+--- a/test/cpp/c10d/CMakeLists.txt
++++ b/test/cpp/c10d/CMakeLists.txt
+@@ -64,7 +64,7 @@ if(USE_CUDA)
+ torch_cpu c10d_cuda_test gtest_main __caffe2_ucc)
+ if(INSTALL_TEST)
+ install(TARGETS ProcessGroupUCCTest DESTINATION bin)
+- install(TARGETS c10d_cuda_test DESTINATION lib)
++ install(TARGETS c10d_cuda_test DESTINATION ${CMAKE_INSTALL_LIBDIR})
+ endif()
+ endif()
+ else()
+--- a/test/cpp/jit/CMakeLists.txt
++++ b/test/cpp/jit/CMakeLists.txt
+@@ -32,9 +32,9 @@ endif()
+ target_link_libraries(backend_with_compiler torch)
+
+ if(INSTALL_TEST)
+- install(TARGETS torchbind_test DESTINATION lib)
+- install(TARGETS jitbackend_test DESTINATION lib)
+- install(TARGETS backend_with_compiler DESTINATION lib)
++ install(TARGETS torchbind_test DESTINATION ${CMAKE_INSTALL_LIBDIR})
++ install(TARGETS jitbackend_test DESTINATION ${CMAKE_INSTALL_LIBDIR})
++ install(TARGETS backend_with_compiler DESTINATION ${CMAKE_INSTALL_LIBDIR})
+ endif()
+
+ # Build the cpp gtest binary containing the cpp-only tests.
diff --git a/sci-libs/caffe2/files/caffe2-2.4.0-rocm-fix-std-cpp17.patch
b/sci-libs/caffe2/files/caffe2-2.4.0-rocm-fix-std-cpp17.patch
new file mode 100644
index 000000000000..3612c3db1a0b
--- /dev/null
+++ b/sci-libs/caffe2/files/caffe2-2.4.0-rocm-fix-std-cpp17.patch
@@ -0,0 +1,50 @@
+Fix for 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
+@@ -1670,6 +1670,7 @@ 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 ${HIP_CXX_FLAGS}) # experiment
++ set_target_properties(torch_hip PROPERTIES CXX_STANDARD 17 CXX_EXTENSIONS
OFF)
+
+ target_link_libraries(torch_hip PUBLIC c10_hip)
+
+@@ -1867,6 +1868,7 @@ if(BUILD_TEST)
+ target_include_directories(${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)
+ install(TARGETS ${test_name} DESTINATION test)
+--- a/cmake/Dependencies.cmake
++++ b/cmake/Dependencies.cmake
+@@ -1050,7 +1050,6 @@ if(USE_ROCM)
+ list(APPEND HIP_CXX_FLAGS -Wno-duplicate-decl-specifier)
+ list(APPEND HIP_CXX_FLAGS -DCAFFE2_USE_MIOPEN)
+ list(APPEND HIP_CXX_FLAGS -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_HIP)
+- list(APPEND HIP_CXX_FLAGS -std=c++17)
+ list(APPEND HIP_CXX_FLAGS -DHIPBLAS_V2)
+ if(HIP_NEW_TYPE_ENUMS)
+ list(APPEND HIP_CXX_FLAGS -DHIP_NEW_TYPE_ENUMS)
+--- a/cmake/public/utils.cmake
++++ b/cmake/public/utils.cmake
+@@ -332,6 +332,7 @@ function(caffe2_hip_binary_target target_name_or_src)
+ caffe2_binary_target(${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()
+
